Add workaround in validation.py for V3 Combo outputs not working as Combo inputs

Updated help text for the --fast argument to clarify potential risks.

comfy/cli_args.py

@@ -105,6 +105,7 @@ cache_group = parser.add_mutually_exclusive_group()
 cache_group.add_argument("--cache-classic", action="store_true", help="Use the old style (aggressive) caching.")
 cache_group.add_argument("--cache-lru", type=int, default=0, help="Use LRU caching with a maximum of N node results cached. May use more RAM/VRAM.")
 cache_group.add_argument("--cache-none", action="store_true", help="Reduced RAM/VRAM usage at the expense of executing every node for each run.")
+cache_group.add_argument("--cache-ram", nargs='?', const=4.0, type=float, default=0, help="Use RAM pressure caching with the specified headroom threshold. If available RAM drops below the threhold the cache remove large items to free RAM. Default 4GB")
 
 attn_group = parser.add_mutually_exclusive_group()
 attn_group.add_argument("--use-split-cross-attention", action="store_true", help="Use the split cross attention optimization. Ignored when xformers is used.")
@@ -132,8 +133,6 @@ parser.add_argument("--reserve-vram", type=float, default=None, help="Set the am
 
 parser.add_argument("--async-offload", action="store_true", help="Use async weight offloading.")
 
-parser.add_argument("--flipflop-offload", action="store_true", help="Use async flipflop weight offloading for supported DiT models.")
-
 parser.add_argument("--force-non-blocking", action="store_true", help="Force ComfyUI to use non-blocking operations for all applicable tensors. This may improve performance on some non-Nvidia systems but can cause issues with some workflows.")
 
 parser.add_argument("--default-hashing-function", type=str, choices=['md5', 'sha1', 'sha256', 'sha512'], default='sha256', help="Allows you to choose the hash function to use for duplicate filename / contents comparison. Default is sha256.")
@@ -146,8 +145,9 @@ class PerformanceFeature(enum.Enum):
     Fp8MatrixMultiplication = "fp8_matrix_mult"
     CublasOps = "cublas_ops"
     AutoTune = "autotune"
+    PinnedMem = "pinned_memory"
 
-parser.add_argument("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. --fast with no arguments enables everything. You can pass a list specific optimizations if you only want to enable specific ones. Current valid optimizations: {}".format(" ".join(map(lambda c: c.value, PerformanceFeature))))
+parser.add_argument("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. This is used to test new features so using it might crash your comfyui. --fast with no arguments enables everything. You can pass a list specific optimizations if you only want to enable specific ones. Current valid optimizations: {}".format(" ".join(map(lambda c: c.value, PerformanceFeature))))
 
 parser.add_argument("--mmap-torch-files", action="store_true", help="Use mmap when loading ckpt/pt files.")
 parser.add_argument("--disable-mmap", action="store_true", help="Don't use mmap when loading safetensors.")

comfy/controlnet.py

@@ -310,11 +310,13 @@ class ControlLoraOps:
             self.bias = None
 
         def forward(self, input):
-            weight, bias = comfy.ops.cast_bias_weight(self, input)
+            weight, bias, offload_stream = comfy.ops.cast_bias_weight(self, input, offloadable=True)
             if self.up is not None:
-                return torch.nn.functional.linear(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias)
+                x = torch.nn.functional.linear(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias)
             else:
-                return torch.nn.functional.linear(input, weight, bias)
+                x = torch.nn.functional.linear(input, weight, bias)
+            comfy.ops.uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
     class Conv2d(torch.nn.Module, comfy.ops.CastWeightBiasOp):
         def __init__(
@@ -350,12 +352,13 @@ class ControlLoraOps:
 
 
         def forward(self, input):
-            weight, bias = comfy.ops.cast_bias_weight(self, input)
+            weight, bias, offload_stream = comfy.ops.cast_bias_weight(self, input, offloadable=True)
             if self.up is not None:
-                return torch.nn.functional.conv2d(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias, self.stride, self.padding, self.dilation, self.groups)
+                x = torch.nn.functional.conv2d(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias, self.stride, self.padding, self.dilation, self.groups)
             else:
-                return torch.nn.functional.conv2d(input, weight, bias, self.stride, self.padding, self.dilation, self.groups)
-
+                x = torch.nn.functional.conv2d(input, weight, bias, self.stride, self.padding, self.dilation, self.groups)
+            comfy.ops.uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
 class ControlLora(ControlNet):
     def __init__(self, control_weights, global_average_pooling=False, model_options={}): #TODO? model_options

comfy/ldm/flipflop_transformer.py

@@ -1,200 +0,0 @@
-from __future__ import annotations
-import torch
-import copy
-
-import comfy.model_management
-
-
-class FlipFlopModule(torch.nn.Module):
-    def __init__(self, block_types: tuple[str, ...], enable_flipflop: bool = True):
-        super().__init__()
-        self.block_types = block_types
-        self.enable_flipflop = enable_flipflop
-        self.flipflop: dict[str, FlipFlopHolder] = {}
-        self.block_info: dict[str, tuple[int, int]] = {}
-        self.flipflop_prefixes: list[str] = []
-
-    def setup_flipflop_holders(self, block_info: dict[str, tuple[int, int]], flipflop_prefixes: list[str], load_device: torch.device, offload_device: torch.device):
-        for block_type, (flipflop_blocks, total_blocks) in block_info.items():
-            if block_type in self.flipflop:
-                continue
-            self.flipflop[block_type] = FlipFlopHolder(getattr(self, block_type)[total_blocks-flipflop_blocks:], flipflop_blocks, total_blocks, load_device, offload_device)
-            self.block_info[block_type] = (flipflop_blocks, total_blocks)
-        self.flipflop_prefixes = flipflop_prefixes.copy()
-
-    def init_flipflop_block_copies(self, device: torch.device) -> int:
-        memory_freed = 0
-        for holder in self.flipflop.values():
-            memory_freed += holder.init_flipflop_block_copies(device)
-        return memory_freed
-
-    def clean_flipflop_holders(self):
-        memory_freed = 0
-        for block_type in list(self.flipflop.keys()):
-            memory_freed += self.flipflop[block_type].clean_flipflop_blocks()
-            del self.flipflop[block_type]
-        self.block_info = {}
-        self.flipflop_prefixes = []
-        return memory_freed
-
-    def get_all_blocks(self, block_type: str) -> list[torch.nn.Module]:
-        return getattr(self, block_type)
-
-    def get_blocks(self, block_type: str) -> torch.nn.ModuleList:
-        if block_type not in self.block_types:
-            raise ValueError(f"Block type {block_type} not found in {self.block_types}")
-        if block_type in self.flipflop:
-            return getattr(self, block_type)[:self.flipflop[block_type].i_offset]
-        return getattr(self, block_type)
-
-    def get_all_block_module_sizes(self, reverse_sort_by_size: bool = False) -> list[tuple[str, int]]:
-        '''
-        Returns a list of (block_type, size) sorted by size.
-        If reverse_sort_by_size is True, the list is sorted by size in reverse order.
-        '''
-        sizes = [(block_type, self.get_block_module_size(block_type)) for block_type in self.block_types]
-        sizes.sort(key=lambda x: x[1], reverse=reverse_sort_by_size)
-        return sizes
-
-    def get_block_module_size(self, block_type: str) -> int:
-        return comfy.model_management.module_size(getattr(self, block_type)[0])
-
-    def execute_blocks(self, block_type: str, func, out: torch.Tensor | tuple[torch.Tensor,...], *args, **kwargs):
-        # execute blocks, supporting both single and double (or higher) block types
-        if isinstance(out, torch.Tensor):
-            out = (out,)
-        for i, block in enumerate(self.get_blocks(block_type)):
-            out = func(i, block, *out, *args, **kwargs)
-            if isinstance(out, torch.Tensor):
-                out = (out,)
-        if block_type in self.flipflop:
-            holder = self.flipflop[block_type]
-            with holder.context() as ctx:
-                for i, block in enumerate(holder.blocks):
-                    out = ctx(func, i, block, *out, *args, **kwargs)
-                    if isinstance(out, torch.Tensor):
-                        out = (out,)
-        if len(out) == 1:
-            out = out[0]
-        return out
-
-
-class FlipFlopContext:
-    def __init__(self, holder: FlipFlopHolder):
-        # NOTE: there is a bug when there are an odd number of blocks to flipflop.
-        # Worked around right now by always making sure it will be even, but need to resolve.
-        self.holder = holder
-        self.reset()
-
-    def reset(self):
-        self.num_blocks = len(self.holder.blocks)
-        self.first_flip = True
-        self.first_flop = True
-        self.last_flip = False
-        self.last_flop = False
-
-    def __enter__(self):
-        self.reset()
-        return self
-
-    def __exit__(self, exc_type, exc_value, traceback):
-        self.holder.compute_stream.record_event(self.holder.cpy_end_event)
-
-    def do_flip(self, func, i: int, _, *args, **kwargs):
-        # flip
-        self.holder.compute_stream.wait_event(self.holder.cpy_end_event)
-        with torch.cuda.stream(self.holder.compute_stream):
-            out = func(i+self.holder.i_offset, self.holder.flip, *args, **kwargs)
-        self.holder.event_flip.record(self.holder.compute_stream)
-        # while flip executes, queue flop to copy to its next block
-        next_flop_i = i + 1
-        if next_flop_i >= self.num_blocks:
-            next_flop_i = next_flop_i - self.num_blocks
-            self.last_flip = True
-        if not self.first_flip:
-            self.holder._copy_state_dict(self.holder.flop.state_dict(), self.holder.blocks[next_flop_i].state_dict(), self.holder.event_flop, self.holder.cpy_end_event)
-        if self.last_flip:
-            self.holder._copy_state_dict(self.holder.flip.state_dict(), self.holder.blocks[0].state_dict(), cpy_start_event=self.holder.event_flip)
-        self.first_flip = False
-        return out
-
-    def do_flop(self, func, i: int, _, *args, **kwargs):
-        # flop
-        if not self.first_flop:
-            self.holder.compute_stream.wait_event(self.holder.cpy_end_event)
-        with torch.cuda.stream(self.holder.compute_stream):
-            out = func(i+self.holder.i_offset, self.holder.flop, *args, **kwargs)
-        self.holder.event_flop.record(self.holder.compute_stream)
-        # while flop executes, queue flip to copy to its next block
-        next_flip_i = i + 1
-        if next_flip_i >= self.num_blocks:
-            next_flip_i = next_flip_i - self.num_blocks
-            self.last_flop = True
-        self.holder._copy_state_dict(self.holder.flip.state_dict(), self.holder.blocks[next_flip_i].state_dict(), self.holder.event_flip, self.holder.cpy_end_event)
-        if self.last_flop:
-            self.holder._copy_state_dict(self.holder.flop.state_dict(), self.holder.blocks[1].state_dict(), cpy_start_event=self.holder.event_flop)
-        self.first_flop = False
-        return out
-
-    @torch.no_grad()
-    def __call__(self, func, i: int, block: torch.nn.Module, *args, **kwargs):
-        # flips are even indexes, flops are odd indexes
-        if i % 2 == 0:
-            return self.do_flip(func, i, block, *args, **kwargs)
-        else:
-            return self.do_flop(func, i, block, *args, **kwargs)
-
-
-class FlipFlopHolder:
-    def __init__(self, blocks: list[torch.nn.Module], flip_amount: int, total_amount: int, load_device: torch.device, offload_device: torch.device):
-        self.load_device = load_device
-        self.offload_device = offload_device
-        self.blocks = blocks
-        self.flip_amount = flip_amount
-        self.total_amount = total_amount
-        # NOTE: used to make sure block indexes passed into block functions match expected patch indexes
-        self.i_offset = total_amount - flip_amount
-
-        self.block_module_size = 0
-        if len(self.blocks) > 0:
-            self.block_module_size = comfy.model_management.module_size(self.blocks[0])
-
-        self.flip: torch.nn.Module = None
-        self.flop: torch.nn.Module = None
-
-        self.compute_stream = torch.cuda.default_stream(self.load_device)
-        self.cpy_stream = torch.cuda.Stream(self.load_device)
-
-        self.event_flip = torch.cuda.Event(enable_timing=False)
-        self.event_flop = torch.cuda.Event(enable_timing=False)
-        self.cpy_end_event = torch.cuda.Event(enable_timing=False)
-        # INIT - is this actually needed?
-        self.compute_stream.record_event(self.cpy_end_event)
-
-    def _copy_state_dict(self, dst, src, cpy_start_event: torch.cuda.Event=None, cpy_end_event: torch.cuda.Event=None):
-        if cpy_start_event:
-            self.cpy_stream.wait_event(cpy_start_event)
-
-        with torch.cuda.stream(self.cpy_stream):
-            for k, v in src.items():
-                dst[k].copy_(v, non_blocking=True)
-        if cpy_end_event:
-            cpy_end_event.record(self.cpy_stream)
-
-    def context(self):
-        return FlipFlopContext(self)
-
-    def init_flipflop_block_copies(self, load_device: torch.device) -> int:
-        self.flip = copy.deepcopy(self.blocks[0]).to(device=load_device)
-        self.flop = copy.deepcopy(self.blocks[1]).to(device=load_device)
-        return comfy.model_management.module_size(self.flip) + comfy.model_management.module_size(self.flop)
-
-    def clean_flipflop_blocks(self) -> int:
-        memory_freed = 0
-        memory_freed += comfy.model_management.module_size(self.flip)
-        memory_freed += comfy.model_management.module_size(self.flop)
-        del self.flip
-        del self.flop
-        self.flip = None
-        self.flop = None
-        return memory_freed

comfy/ldm/flux/model.py

@@ -7,7 +7,6 @@ from torch import Tensor, nn
 from einops import rearrange, repeat
 import comfy.ldm.common_dit
 import comfy.patcher_extension
-from comfy.ldm.flipflop_transformer import FlipFlopModule
 
 from .layers import (
     DoubleStreamBlock,
@@ -36,13 +35,13 @@ class FluxParams:
     guidance_embed: bool
 
 
-class Flux(FlipFlopModule):
+class Flux(nn.Module):
     """
     Transformer model for flow matching on sequences.
     """
 
     def __init__(self, image_model=None, final_layer=True, dtype=None, device=None, operations=None, **kwargs):
-        super().__init__(("double_blocks", "single_blocks"))
+        super().__init__()
         self.dtype = dtype
         params = FluxParams(**kwargs)
         self.params = params
@@ -90,72 +89,6 @@ class Flux(FlipFlopModule):
         if final_layer:
             self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels, dtype=dtype, device=device, operations=operations)
 
-    def indiv_double_block_fwd(self, i, block, img, txt, vec, pe, attn_mask, control, blocks_replace, transformer_options):
-        if ("double_block", i) in blocks_replace:
-            def block_wrap(args):
-                out = {}
-                out["img"], out["txt"] = block(img=args["img"],
-                                               txt=args["txt"],
-                                               vec=args["vec"],
-                                               pe=args["pe"],
-                                               attn_mask=args.get("attn_mask"),
-                                               transformer_options=args.get("transformer_options"))
-                return out
-
-            out = blocks_replace[("double_block", i)]({"img": img,
-                                                       "txt": txt,
-                                                       "vec": vec,
-                                                       "pe": pe,
-                                                       "attn_mask": attn_mask,
-                                                       "transformer_options": transformer_options},
-                                                       {"original_block": block_wrap})
-            txt = out["txt"]
-            img = out["img"]
-        else:
-            img, txt = block(img=img,
-                             txt=txt,
-                             vec=vec,
-                             pe=pe,
-                             attn_mask=attn_mask,
-                             transformer_options=transformer_options)
-
-        if control is not None: # Controlnet
-            control_i = control.get("input")
-            if i < len(control_i):
-                add = control_i[i]
-                if add is not None:
-                    img[:, :add.shape[1]] += add
-        return img, txt
-
-    def indiv_single_block_fwd(self, i, block, img, txt, vec, pe, attn_mask, control, blocks_replace, transformer_options):
-        if ("single_block", i) in blocks_replace:
-            def block_wrap(args):
-                out = {}
-                out["img"] = block(args["img"],
-                                   vec=args["vec"],
-                                   pe=args["pe"],
-                                   attn_mask=args.get("attn_mask"),
-                                   transformer_options=args.get("transformer_options"))
-                return out
-
-            out = blocks_replace[("single_block", i)]({"img": img,
-                                                       "vec": vec,
-                                                       "pe": pe,
-                                                       "attn_mask": attn_mask,
-                                                       "transformer_options": transformer_options},
-                                                       {"original_block": block_wrap})
-            img = out["img"]
-        else:
-            img = block(img, vec=vec, pe=pe, attn_mask=attn_mask, transformer_options=transformer_options)
-
-        if control is not None: # Controlnet
-            control_o = control.get("output")
-            if i < len(control_o):
-                add = control_o[i]
-                if add is not None:
-                    img[:, txt.shape[1] : txt.shape[1] + add.shape[1], ...] += add
-        return img
-
     def forward_orig(
         self,
         img: Tensor,
@@ -203,16 +136,74 @@ class Flux(FlipFlopModule):
             pe = None
 
         blocks_replace = patches_replace.get("dit", {})
-        # execute double blocks
-        img, txt = self.execute_blocks("double_blocks", self.indiv_double_block_fwd, (img, txt), vec, pe, attn_mask, control, blocks_replace, transformer_options)
+        for i, block in enumerate(self.double_blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"], out["txt"] = block(img=args["img"],
+                                                   txt=args["txt"],
+                                                   vec=args["vec"],
+                                                   pe=args["pe"],
+                                                   attn_mask=args.get("attn_mask"),
+                                                   transformer_options=args.get("transformer_options"))
+                    return out
+
+                out = blocks_replace[("double_block", i)]({"img": img,
+                                                           "txt": txt,
+                                                           "vec": vec,
+                                                           "pe": pe,
+                                                           "attn_mask": attn_mask,
+                                                           "transformer_options": transformer_options},
+                                                          {"original_block": block_wrap})
+                txt = out["txt"]
+                img = out["img"]
+            else:
+                img, txt = block(img=img,
+                                 txt=txt,
+                                 vec=vec,
+                                 pe=pe,
+                                 attn_mask=attn_mask,
+                                 transformer_options=transformer_options)
+
+            if control is not None: # Controlnet
+                control_i = control.get("input")
+                if i < len(control_i):
+                    add = control_i[i]
+                    if add is not None:
+                        img[:, :add.shape[1]] += add
 
         if img.dtype == torch.float16:
             img = torch.nan_to_num(img, nan=0.0, posinf=65504, neginf=-65504)
 
         img = torch.cat((txt, img), 1)
 
-        # execute single blocks
-        img = self.execute_blocks("single_blocks", self.indiv_single_block_fwd, img, txt, vec, pe, attn_mask, control, blocks_replace, transformer_options)
+        for i, block in enumerate(self.single_blocks):
+            if ("single_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(args["img"],
+                                       vec=args["vec"],
+                                       pe=args["pe"],
+                                       attn_mask=args.get("attn_mask"),
+                                       transformer_options=args.get("transformer_options"))
+                    return out
+
+                out = blocks_replace[("single_block", i)]({"img": img,
+                                                           "vec": vec,
+                                                           "pe": pe,
+                                                           "attn_mask": attn_mask,
+                                                           "transformer_options": transformer_options},
+                                                          {"original_block": block_wrap})
+                img = out["img"]
+            else:
+                img = block(img, vec=vec, pe=pe, attn_mask=attn_mask, transformer_options=transformer_options)
+
+            if control is not None: # Controlnet
+                control_o = control.get("output")
+                if i < len(control_o):
+                    add = control_o[i]
+                    if add is not None:
+                        img[:, txt.shape[1] : txt.shape[1] + add.shape[1], ...] += add
 
         img = img[:, txt.shape[1] :, ...]
 

comfy/ldm/lumina/model.py

@@ -522,7 +522,7 @@ class NextDiT(nn.Module):
         max_cap_len = max(l_effective_cap_len)
         max_img_len = max(l_effective_img_len)
 
-        position_ids = torch.zeros(bsz, max_seq_len, 3, dtype=torch.int32, device=device)
+        position_ids = torch.zeros(bsz, max_seq_len, 3, dtype=torch.float32, device=device)
 
         for i in range(bsz):
             cap_len = l_effective_cap_len[i]
@@ -531,10 +531,22 @@ class NextDiT(nn.Module):
             H_tokens, W_tokens = H // pH, W // pW
             assert H_tokens * W_tokens == img_len
 
-            position_ids[i, :cap_len, 0] = torch.arange(cap_len, dtype=torch.int32, device=device)
+            rope_options = transformer_options.get("rope_options", None)
+            h_scale = 1.0
+            w_scale = 1.0
+            h_start = 0
+            w_start = 0
+            if rope_options is not None:
+                h_scale = rope_options.get("scale_y", 1.0)
+                w_scale = rope_options.get("scale_x", 1.0)
+
+                h_start = rope_options.get("shift_y", 0.0)
+                w_start = rope_options.get("shift_x", 0.0)
+
+            position_ids[i, :cap_len, 0] = torch.arange(cap_len, dtype=torch.float32, device=device)
             position_ids[i, cap_len:cap_len+img_len, 0] = cap_len
-            row_ids = torch.arange(H_tokens, dtype=torch.int32, device=device).view(-1, 1).repeat(1, W_tokens).flatten()
-            col_ids = torch.arange(W_tokens, dtype=torch.int32, device=device).view(1, -1).repeat(H_tokens, 1).flatten()
+            row_ids = (torch.arange(H_tokens, dtype=torch.float32, device=device) * h_scale + h_start).view(-1, 1).repeat(1, W_tokens).flatten()
+            col_ids = (torch.arange(W_tokens, dtype=torch.float32, device=device) * w_scale + w_start).view(1, -1).repeat(H_tokens, 1).flatten()
             position_ids[i, cap_len:cap_len+img_len, 1] = row_ids
             position_ids[i, cap_len:cap_len+img_len, 2] = col_ids
 

comfy/ldm/qwen_image/model.py

@@ -5,13 +5,11 @@ import torch.nn.functional as F
 from typing import Optional, Tuple
 from einops import repeat
 
-from comfy.ldm.flipflop_transformer import FlipFlopModule
 from comfy.ldm.lightricks.model import TimestepEmbedding, Timesteps
 from comfy.ldm.modules.attention import optimized_attention_masked
 from comfy.ldm.flux.layers import EmbedND
 import comfy.ldm.common_dit
 import comfy.patcher_extension
-import comfy.ops
 
 class GELU(nn.Module):
     def __init__(self, dim_in: int, dim_out: int, approximate: str = "none", bias: bool = True, dtype=None, device=None, operations=None):
@@ -285,7 +283,7 @@ class LastLayer(nn.Module):
         return x
 
 
-class QwenImageTransformer2DModel(FlipFlopModule):
+class QwenImageTransformer2DModel(nn.Module):
     def __init__(
         self,
         patch_size: int = 2,
@@ -302,9 +300,9 @@ class QwenImageTransformer2DModel(FlipFlopModule):
         final_layer=True,
         dtype=None,
         device=None,
-        operations: comfy.ops.disable_weight_init=None,
+        operations=None,
     ):
-        super().__init__(block_types=("transformer_blocks",))
+        super().__init__()
         self.dtype = dtype
         self.patch_size = patch_size
         self.in_channels = in_channels
@@ -368,40 +366,6 @@ class QwenImageTransformer2DModel(FlipFlopModule):
             comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
         ).execute(x, timestep, context, attention_mask, guidance, ref_latents, transformer_options, **kwargs)
 
-    def indiv_block_fwd(self, i, block, hidden_states, encoder_hidden_states, encoder_hidden_states_mask, temb, image_rotary_emb, patches, control, blocks_replace, x, transformer_options):
-        if ("double_block", i) in blocks_replace:
-            def block_wrap(args):
-                out = {}
-                out["txt"], out["img"] = block(hidden_states=args["img"], encoder_hidden_states=args["txt"], encoder_hidden_states_mask=encoder_hidden_states_mask, temb=args["vec"], image_rotary_emb=args["pe"], transformer_options=args["transformer_options"])
-                return out
-            out = blocks_replace[("double_block", i)]({"img": hidden_states, "txt": encoder_hidden_states, "vec": temb, "pe": image_rotary_emb, "transformer_options": transformer_options}, {"original_block": block_wrap})
-            hidden_states = out["img"]
-            encoder_hidden_states = out["txt"]
-        else:
-            encoder_hidden_states, hidden_states = block(
-                hidden_states=hidden_states,
-                encoder_hidden_states=encoder_hidden_states,
-                encoder_hidden_states_mask=encoder_hidden_states_mask,
-                temb=temb,
-                image_rotary_emb=image_rotary_emb,
-                transformer_options=transformer_options,
-            )
-
-        if "double_block" in patches:
-            for p in patches["double_block"]:
-                out = p({"img": hidden_states, "txt": encoder_hidden_states, "x": x, "block_index": i, "transformer_options": transformer_options})
-                hidden_states = out["img"]
-                encoder_hidden_states = out["txt"]
-
-        if control is not None: # Controlnet
-            control_i = control.get("input")
-            if i < len(control_i):
-                add = control_i[i]
-                if add is not None:
-                    hidden_states[:, :add.shape[1]] += add
-
-        return hidden_states, encoder_hidden_states
-
     def _forward(
         self,
         x,
@@ -469,8 +433,37 @@ class QwenImageTransformer2DModel(FlipFlopModule):
         patches = transformer_options.get("patches", {})
         blocks_replace = patches_replace.get("dit", {})
 
-        out = (hidden_states, encoder_hidden_states)
-        hidden_states, encoder_hidden_states = self.execute_blocks("transformer_blocks", self.indiv_block_fwd, out, encoder_hidden_states_mask, temb, image_rotary_emb, patches, control, blocks_replace, x, transformer_options)
+        for i, block in enumerate(self.transformer_blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["txt"], out["img"] = block(hidden_states=args["img"], encoder_hidden_states=args["txt"], encoder_hidden_states_mask=encoder_hidden_states_mask, temb=args["vec"], image_rotary_emb=args["pe"], transformer_options=args["transformer_options"])
+                    return out
+                out = blocks_replace[("double_block", i)]({"img": hidden_states, "txt": encoder_hidden_states, "vec": temb, "pe": image_rotary_emb, "transformer_options": transformer_options}, {"original_block": block_wrap})
+                hidden_states = out["img"]
+                encoder_hidden_states = out["txt"]
+            else:
+                encoder_hidden_states, hidden_states = block(
+                    hidden_states=hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_hidden_states_mask=encoder_hidden_states_mask,
+                    temb=temb,
+                    image_rotary_emb=image_rotary_emb,
+                    transformer_options=transformer_options,
+                )
+
+            if "double_block" in patches:
+                for p in patches["double_block"]:
+                    out = p({"img": hidden_states, "txt": encoder_hidden_states, "x": x, "block_index": i, "transformer_options": transformer_options})
+                    hidden_states = out["img"]
+                    encoder_hidden_states = out["txt"]
+
+            if control is not None: # Controlnet
+                control_i = control.get("input")
+                if i < len(control_i):
+                    add = control_i[i]
+                    if add is not None:
+                        hidden_states[:, :add.shape[1]] += add
 
         hidden_states = self.norm_out(hidden_states, temb)
         hidden_states = self.proj_out(hidden_states)

comfy/ldm/wan/model.py

@@ -7,7 +7,6 @@ import torch.nn as nn
 from einops import rearrange
 
 from comfy.ldm.modules.attention import optimized_attention
-from comfy.ldm.flipflop_transformer import FlipFlopModule
 from comfy.ldm.flux.layers import EmbedND
 from comfy.ldm.flux.math import apply_rope1
 import comfy.ldm.common_dit
@@ -385,7 +384,7 @@ class MLPProj(torch.nn.Module):
         return clip_extra_context_tokens
 
 
-class WanModel(FlipFlopModule):
+class WanModel(torch.nn.Module):
     r"""
     Wan diffusion backbone supporting both text-to-video and image-to-video.
     """
@@ -413,7 +412,6 @@ class WanModel(FlipFlopModule):
                  device=None,
                  dtype=None,
                  operations=None,
-                 enable_flipflop=True,
                  ):
         r"""
         Initialize the diffusion model backbone.
@@ -451,7 +449,7 @@ class WanModel(FlipFlopModule):
                 Epsilon value for normalization layers
         """
 
-        super().__init__(block_types=("blocks",), enable_flipflop=enable_flipflop)
+        super().__init__()
         self.dtype = dtype
         operation_settings = {"operations": operations, "device": device, "dtype": dtype}
 
@@ -508,18 +506,6 @@ class WanModel(FlipFlopModule):
         else:
             self.ref_conv = None
 
-    def indiv_block_fwd(self, i, block, x, e0, freqs, context, context_img_len, blocks_replace, transformer_options):
-        if ("double_block", i) in blocks_replace:
-            def block_wrap(args):
-                out = {}
-                out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], context_img_len=context_img_len, transformer_options=args["transformer_options"])
-                return out
-            out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs, "transformer_options": transformer_options}, {"original_block": block_wrap})
-            x = out["img"]
-        else:
-            x = block(x, e=e0, freqs=freqs, context=context, context_img_len=context_img_len, transformer_options=transformer_options)
-        return x
-
     def forward_orig(
         self,
         x,
@@ -581,8 +567,16 @@ class WanModel(FlipFlopModule):
 
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
-        # execute blocks
-        x = self.execute_blocks("blocks", self.indiv_block_fwd, x, e0, freqs, context, context_img_len, blocks_replace, transformer_options)
+        for i, block in enumerate(self.blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], context_img_len=context_img_len, transformer_options=args["transformer_options"])
+                    return out
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs, "transformer_options": transformer_options}, {"original_block": block_wrap})
+                x = out["img"]
+            else:
+                x = block(x, e=e0, freqs=freqs, context=context, context_img_len=context_img_len, transformer_options=transformer_options)
 
         # head
         x = self.head(x, e)
@@ -594,7 +588,7 @@ class WanModel(FlipFlopModule):
         x = self.unpatchify(x, grid_sizes)
         return x
 
-    def rope_encode(self, t, h, w, t_start=0, steps_t=None, steps_h=None, steps_w=None, device=None, dtype=None):
+    def rope_encode(self, t, h, w, t_start=0, steps_t=None, steps_h=None, steps_w=None, device=None, dtype=None, transformer_options={}):
         patch_size = self.patch_size
         t_len = ((t + (patch_size[0] // 2)) // patch_size[0])
         h_len = ((h + (patch_size[1] // 2)) // patch_size[1])
@@ -607,10 +601,22 @@ class WanModel(FlipFlopModule):
         if steps_w is None:
             steps_w = w_len
 
+        h_start = 0
+        w_start = 0
+        rope_options = transformer_options.get("rope_options", None)
+        if rope_options is not None:
+            t_len = (t_len - 1.0) * rope_options.get("scale_t", 1.0) + 1.0
+            h_len = (h_len - 1.0) * rope_options.get("scale_y", 1.0) + 1.0
+            w_len = (w_len - 1.0) * rope_options.get("scale_x", 1.0) + 1.0
+
+            t_start += rope_options.get("shift_t", 0.0)
+            h_start += rope_options.get("shift_y", 0.0)
+            w_start += rope_options.get("shift_x", 0.0)
+
         img_ids = torch.zeros((steps_t, steps_h, steps_w, 3), device=device, dtype=dtype)
         img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(t_start, t_start + (t_len - 1), steps=steps_t, device=device, dtype=dtype).reshape(-1, 1, 1)
-        img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(0, h_len - 1, steps=steps_h, device=device, dtype=dtype).reshape(1, -1, 1)
-        img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(0, w_len - 1, steps=steps_w, device=device, dtype=dtype).reshape(1, 1, -1)
+        img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(h_start, h_start + (h_len - 1), steps=steps_h, device=device, dtype=dtype).reshape(1, -1, 1)
+        img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(w_start, w_start + (w_len - 1), steps=steps_w, device=device, dtype=dtype).reshape(1, 1, -1)
         img_ids = img_ids.reshape(1, -1, img_ids.shape[-1])
 
         freqs = self.rope_embedder(img_ids).movedim(1, 2)
@@ -636,7 +642,7 @@ class WanModel(FlipFlopModule):
         if self.ref_conv is not None and "reference_latent" in kwargs:
             t_len += 1
 
-        freqs = self.rope_encode(t_len, h, w, device=x.device, dtype=x.dtype)
+        freqs = self.rope_encode(t_len, h, w, device=x.device, dtype=x.dtype, transformer_options=transformer_options)
         return self.forward_orig(x, timestep, context, clip_fea=clip_fea, freqs=freqs, transformer_options=transformer_options, **kwargs)[:, :, :t, :h, :w]
 
     def unpatchify(self, x, grid_sizes):
@@ -694,7 +700,7 @@ class VaceWanModel(WanModel):
                  operations=None,
                  ):
 
-        super().__init__(model_type='t2v', patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, flf_pos_embed_token_number=flf_pos_embed_token_number, image_model=image_model, device=device, dtype=dtype, operations=operations, enable_flipflop=False)
+        super().__init__(model_type='t2v', patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, flf_pos_embed_token_number=flf_pos_embed_token_number, image_model=image_model, device=device, dtype=dtype, operations=operations)
         operation_settings = {"operations": operations, "device": device, "dtype": dtype}
 
         # Vace
@@ -814,7 +820,7 @@ class CameraWanModel(WanModel):
         else:
             model_type = 't2v'
 
-        super().__init__(model_type=model_type, patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, flf_pos_embed_token_number=flf_pos_embed_token_number, image_model=image_model, device=device, dtype=dtype, operations=operations, enable_flipflop=False)
+        super().__init__(model_type=model_type, patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, flf_pos_embed_token_number=flf_pos_embed_token_number, image_model=image_model, device=device, dtype=dtype, operations=operations)
         operation_settings = {"operations": operations, "device": device, "dtype": dtype}
 
         self.control_adapter = WanCamAdapter(in_dim_control_adapter, dim, kernel_size=patch_size[1:], stride=patch_size[1:], operation_settings=operation_settings)
@@ -1217,7 +1223,7 @@ class WanModel_S2V(WanModel):
                  operations=None,
                  ):
 
-        super().__init__(model_type='t2v', patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, image_model=image_model, device=device, dtype=dtype, operations=operations, enable_flipflop=False)
+        super().__init__(model_type='t2v', patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, image_model=image_model, device=device, dtype=dtype, operations=operations)
 
         self.trainable_cond_mask = operations.Embedding(3, self.dim, device=device, dtype=dtype)
 
@@ -1517,7 +1523,7 @@ class HumoWanModel(WanModel):
                  operations=None,
                  ):
 
-        super().__init__(model_type='t2v', patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, flf_pos_embed_token_number=flf_pos_embed_token_number, wan_attn_block_class=WanAttentionBlockAudio, image_model=image_model, device=device, dtype=dtype, operations=operations, enable_flipflop=False)
+        super().__init__(model_type='t2v', patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, flf_pos_embed_token_number=flf_pos_embed_token_number, wan_attn_block_class=WanAttentionBlockAudio, image_model=image_model, device=device, dtype=dtype, operations=operations)
 
         self.audio_proj = AudioProjModel(seq_len=8, blocks=5, channels=1280, intermediate_dim=512, output_dim=1536, context_tokens=audio_token_num, dtype=dtype, device=device, operations=operations)
 

comfy/ldm/wan/model_animate.py

@@ -426,7 +426,7 @@ class AnimateWanModel(WanModel):
                  operations=None,
                  ):
 
-        super().__init__(model_type='i2v', patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, flf_pos_embed_token_number=flf_pos_embed_token_number, image_model=image_model, device=device, dtype=dtype, operations=operations, enable_flipflop=False)
+        super().__init__(model_type='i2v', patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, flf_pos_embed_token_number=flf_pos_embed_token_number, image_model=image_model, device=device, dtype=dtype, operations=operations)
 
         self.pose_patch_embedding = operations.Conv3d(
             16, dim, kernel_size=patch_size, stride=patch_size, device=device, dtype=dtype

comfy/model_management.py

@@ -1006,8 +1006,6 @@ def force_channels_last():
     #TODO
     return False
 
-def flipflop_enabled():
-    return args.flipflop_offload
 
 STREAMS = {}
 NUM_STREAMS = 1
@@ -1015,6 +1013,16 @@ if args.async_offload:
     NUM_STREAMS = 2
     logging.info("Using async weight offloading with {} streams".format(NUM_STREAMS))
 
+def current_stream(device):
+    if device is None:
+        return None
+    if is_device_cuda(device):
+        return torch.cuda.current_stream()
+    elif is_device_xpu(device):
+        return torch.xpu.current_stream()
+    else:
+        return None
+
 stream_counters = {}
 def get_offload_stream(device):
     stream_counter = stream_counters.get(device, 0)
@@ -1023,21 +1031,17 @@ def get_offload_stream(device):
 
     if device in STREAMS:
         ss = STREAMS[device]
-        s = ss[stream_counter]
+        #Sync the oldest stream in the queue with the current
+        ss[stream_counter].wait_stream(current_stream(device))
         stream_counter = (stream_counter + 1) % len(ss)
-        if is_device_cuda(device):
-            ss[stream_counter].wait_stream(torch.cuda.current_stream())
-        elif is_device_xpu(device):
-            ss[stream_counter].wait_stream(torch.xpu.current_stream())
         stream_counters[device] = stream_counter
-        return s
+        return ss[stream_counter]
     elif is_device_cuda(device):
         ss = []
         for k in range(NUM_STREAMS):
             ss.append(torch.cuda.Stream(device=device, priority=0))
         STREAMS[device] = ss
         s = ss[stream_counter]
-        stream_counter = (stream_counter + 1) % len(ss)
         stream_counters[device] = stream_counter
         return s
     elif is_device_xpu(device):
@@ -1046,18 +1050,14 @@ def get_offload_stream(device):
             ss.append(torch.xpu.Stream(device=device, priority=0))
         STREAMS[device] = ss
         s = ss[stream_counter]
-        stream_counter = (stream_counter + 1) % len(ss)
         stream_counters[device] = stream_counter
         return s
     return None
 
 def sync_stream(device, stream):
-    if stream is None:
+    if stream is None or current_stream(device) is None:
         return
-    if is_device_cuda(device):
-        torch.cuda.current_stream().wait_stream(stream)
-    elif is_device_xpu(device):
-        torch.xpu.current_stream().wait_stream(stream)
+    current_stream(device).wait_stream(stream)
 
 def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, stream=None):
     if device is None or weight.device == device:
@@ -1082,6 +1082,36 @@ def cast_to_device(tensor, device, dtype, copy=False):
     non_blocking = device_supports_non_blocking(device)
     return cast_to(tensor, dtype=dtype, device=device, non_blocking=non_blocking, copy=copy)
 
+def pin_memory(tensor):
+    if PerformanceFeature.PinnedMem not in args.fast:
+        return False
+
+    if not is_nvidia():
+        return False
+
+    if not is_device_cpu(tensor.device):
+        return False
+
+    if torch.cuda.cudart().cudaHostRegister(tensor.data_ptr(), tensor.numel() * tensor.element_size(), 1) == 0:
+        return True
+
+    return False
+
+def unpin_memory(tensor):
+    if PerformanceFeature.PinnedMem not in args.fast:
+        return False
+
+    if not is_nvidia():
+        return False
+
+    if not is_device_cpu(tensor.device):
+        return False
+
+    if torch.cuda.cudart().cudaHostUnregister(tensor.data_ptr()) == 0:
+        return True
+
+    return False
+
 def sage_attention_enabled():
     return args.use_sage_attention
 

comfy/model_patcher.py

@@ -25,7 +25,7 @@ import logging
 import math
 import uuid
 from typing import Callable, Optional
-import time # TODO remove
+
 import torch
 
 import comfy.float
@@ -238,6 +238,7 @@ class ModelPatcher:
         self.force_cast_weights = False
         self.patches_uuid = uuid.uuid4()
         self.parent = None
+        self.pinned = set()
 
         self.attachments: dict[str] = {}
         self.additional_models: dict[str, list[ModelPatcher]] = {}
@@ -275,6 +276,9 @@ class ModelPatcher:
         self.size = comfy.model_management.module_size(self.model)
         return self.size
 
+    def get_ram_usage(self):
+        return self.model_size()
+
     def loaded_size(self):
         return self.model.model_loaded_weight_memory
 
@@ -294,6 +298,7 @@ class ModelPatcher:
         n.backup = self.backup
         n.object_patches_backup = self.object_patches_backup
         n.parent = self
+        n.pinned = self.pinned
 
         n.force_cast_weights = self.force_cast_weights
 
@@ -450,6 +455,19 @@ class ModelPatcher:
     def set_model_post_input_patch(self, patch):
         self.set_model_patch(patch, "post_input")
 
+    def set_model_rope_options(self, scale_x, shift_x, scale_y, shift_y, scale_t, shift_t, **kwargs):
+        rope_options = self.model_options["transformer_options"].get("rope_options", {})
+        rope_options["scale_x"] = scale_x
+        rope_options["scale_y"] = scale_y
+        rope_options["scale_t"] = scale_t
+
+        rope_options["shift_x"] = shift_x
+        rope_options["shift_y"] = shift_y
+        rope_options["shift_t"] = shift_t
+
+        self.model_options["transformer_options"]["rope_options"] = rope_options
+
+
     def add_object_patch(self, name, obj):
         self.object_patches[name] = obj
 
@@ -591,7 +609,7 @@ class ModelPatcher:
                         sd.pop(k)
             return sd
 
-    def patch_weight_to_device(self, key, device_to=None, inplace_update=False, device_final=None):
+    def patch_weight_to_device(self, key, device_to=None, inplace_update=False):
         if key not in self.patches:
             return
 
@@ -611,103 +629,30 @@ class ModelPatcher:
         out_weight = comfy.lora.calculate_weight(self.patches[key], temp_weight, key)
         if set_func is None:
             out_weight = comfy.float.stochastic_rounding(out_weight, weight.dtype, seed=string_to_seed(key))
-            if device_final is not None:
-                out_weight = out_weight.to(device_final)
             if inplace_update:
                 comfy.utils.copy_to_param(self.model, key, out_weight)
             else:
                 comfy.utils.set_attr_param(self.model, key, out_weight)
         else:
-            if device_final is not None:
-                out_weight = out_weight.to(device_final)
             set_func(out_weight, inplace_update=inplace_update, seed=string_to_seed(key))
 
-    def supports_flipflop(self):
-        # flipflop requires diffusion_model, explicit flipflop support, NVIDIA CUDA streams, and loading/offloading VRAM
-        if not comfy.model_management.flipflop_enabled():
-            return False
-        if not hasattr(self.model, "diffusion_model"):
-            return False
-        if not getattr(self.model.diffusion_model, "enable_flipflop", False):
-            return False
-        if not comfy.model_management.is_nvidia():
-            return False
-        if comfy.model_management.vram_state in (comfy.model_management.VRAMState.HIGH_VRAM, comfy.model_management.VRAMState.SHARED):
-            return False
-        return True
+    def pin_weight_to_device(self, key):
+        weight, set_func, convert_func = get_key_weight(self.model, key)
+        if comfy.model_management.pin_memory(weight):
+            self.pinned.add(key)
 
-    def setup_flipflop(self, flipflop_blocks_per_type: dict[str, tuple[int, int]], flipflop_prefixes: list[str]):
-        if not self.supports_flipflop():
-            return
-        logging.info(f"setting up flipflop with {flipflop_blocks_per_type}")
-        self.model.diffusion_model.setup_flipflop_holders(flipflop_blocks_per_type, flipflop_prefixes, self.load_device, self.offload_device)
+    def unpin_weight(self, key):
+        if key in self.pinned:
+            weight, set_func, convert_func = get_key_weight(self.model, key)
+            comfy.model_management.unpin_memory(weight)
+            self.pinned.remove(key)
 
-    def init_flipflop_block_copies(self) -> int:
-        if not self.supports_flipflop():
-            return 0
-        return self.model.diffusion_model.init_flipflop_block_copies(self.load_device)
+    def unpin_all_weights(self):
+        for key in list(self.pinned):
+            self.unpin_weight(key)
 
-    def clean_flipflop(self) -> int:
-        if not self.supports_flipflop():
-            return 0
-        return self.model.diffusion_model.clean_flipflop_holders()
-
-    def _get_existing_flipflop_prefixes(self):
-        if self.supports_flipflop():
-            return self.model.diffusion_model.flipflop_prefixes
-        return []
-
-    def _calc_flipflop_prefixes(self, lowvram_model_memory=0, prepare_flipflop=False):
-        flipflop_prefixes = []
-        flipflop_blocks_per_type: dict[str, tuple[int, int]] = {}
-        if lowvram_model_memory > 0 and self.supports_flipflop():
-            block_buffer = 3
-            valid_block_types = []
-            # for each block type, check if have enough room to flipflop
-            for block_info in self.model.diffusion_model.get_all_block_module_sizes(reverse_sort_by_size=True):
-                block_size: int = block_info[1]
-                if block_size * block_buffer < lowvram_model_memory:
-                    valid_block_types.append(block_info)
-            # if have candidates for flipping, see how many of each type we have can flipflop
-            if len(valid_block_types) > 0:
-                leftover_memory = lowvram_model_memory
-                for block_info in valid_block_types:
-                    block_type: str = block_info[0]
-                    block_size: int = block_info[1]
-                    total_blocks = len(self.model.diffusion_model.get_all_blocks(block_type))
-                    n_fit_in_memory = int(leftover_memory // block_size)
-                    # if all (or more) of this block type would fit in memory, no need to flipflop with it
-                    if n_fit_in_memory >= total_blocks:
-                        leftover_memory -= total_blocks * block_size
-                        continue
-                    # if the amount of this block that would fit in memory is less than buffer, skip this block type
-                    if n_fit_in_memory < block_buffer:
-                        continue
-                    # 2 blocks worth of VRAM may be needed for flipflop, so make sure to account for them.
-                    flipflop_blocks = min((total_blocks - n_fit_in_memory) + 2, total_blocks)
-                    # for now, work around odd number issue by making it even
-                    if flipflop_blocks % 2 != 0:
-                        if flipflop_blocks == total_blocks:
-                            flipflop_blocks -= 1
-                        else:
-                            flipflop_blocks += 1
-                    flipflop_blocks_per_type[block_type] = (flipflop_blocks, total_blocks)
-                    leftover_memory -= (total_blocks - flipflop_blocks + 2) * block_size
-                # if there are blocks to flipflop, need to mark their keys
-                for block_type, (flipflop_blocks, total_blocks) in flipflop_blocks_per_type.items():
-                    # blocks to flipflop are at the end
-                    for i in range(total_blocks-flipflop_blocks, total_blocks):
-                        flipflop_prefixes.append(f"diffusion_model.{block_type}.{i}")
-        if prepare_flipflop and len(flipflop_blocks_per_type) > 0:
-            self.setup_flipflop(flipflop_blocks_per_type, flipflop_prefixes)
-        return flipflop_prefixes
-
-    def _load_list(self, lowvram_model_memory=0, prepare_flipflop=False, get_existing_flipflop=False):
+    def _load_list(self):
         loading = []
-        if get_existing_flipflop:
-            flipflop_prefixes = self._get_existing_flipflop_prefixes()
-        else:
-            flipflop_prefixes = self._calc_flipflop_prefixes(lowvram_model_memory, prepare_flipflop)
         for n, m in self.model.named_modules():
             params = []
             skip = False
@@ -718,12 +663,7 @@ class ModelPatcher:
                     skip = True # skip random weights in non leaf modules
                     break
             if not skip and (hasattr(m, "comfy_cast_weights") or len(params) > 0):
-                flipflop = False
-                for prefix in flipflop_prefixes:
-                    if n.startswith(prefix):
-                        flipflop = True
-                        break
-                loading.append((comfy.model_management.module_size(m), n, m, params, flipflop))
+                loading.append((comfy.model_management.module_size(m), n, m, params))
         return loading
 
     def load(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False, full_load=False):
@@ -733,18 +673,15 @@ class ModelPatcher:
             patch_counter = 0
             lowvram_counter = 0
             lowvram_mem_counter = 0
-            flipflop_counter = 0
-            flipflop_mem_counter = 0
-            loading = self._load_list(lowvram_model_memory, prepare_flipflop=True)
+            loading = self._load_list()
 
             load_completely = []
-            load_flipflop = []
+            offloaded = []
             loading.sort(reverse=True)
             for x in loading:
                 n = x[1]
                 m = x[2]
                 params = x[3]
-                flipflop: bool = x[4]
                 module_mem = x[0]
 
                 lowvram_weight = False
@@ -752,7 +689,7 @@ class ModelPatcher:
                 weight_key = "{}.weight".format(n)
                 bias_key = "{}.bias".format(n)
 
-                if not full_load and hasattr(m, "comfy_cast_weights") and not flipflop:
+                if not full_load and hasattr(m, "comfy_cast_weights"):
                     if mem_counter + module_mem >= lowvram_model_memory:
                         lowvram_weight = True
                         lowvram_counter += 1
@@ -782,15 +719,12 @@ class ModelPatcher:
                             patch_counter += 1
 
                     cast_weight = True
+                    offloaded.append((module_mem, n, m, params))
                 else:
                     if hasattr(m, "comfy_cast_weights"):
                         wipe_lowvram_weight(m)
 
-                    if flipflop:
-                        flipflop_counter += 1
-                        flipflop_mem_counter += module_mem
-                        load_flipflop.append((module_mem, n, m, params))
-                    elif full_load or mem_counter + module_mem < lowvram_model_memory:
+                    if full_load or mem_counter + module_mem < lowvram_model_memory:
                         mem_counter += module_mem
                         load_completely.append((module_mem, n, m, params))
 
@@ -806,7 +740,6 @@ class ModelPatcher:
 
                 mem_counter += move_weight_functions(m, device_to)
 
-            # handle load completely
             load_completely.sort(reverse=True)
             for x in load_completely:
                 n = x[1]
@@ -817,7 +750,9 @@ class ModelPatcher:
                         continue
 
                 for param in params:
-                    self.patch_weight_to_device("{}.{}".format(n, param), device_to=device_to)
+                    key = "{}.{}".format(n, param)
+                    self.unpin_weight(key)
+                    self.patch_weight_to_device(key, device_to=device_to)
 
                 logging.debug("lowvram: loaded module regularly {} {}".format(n, m))
                 m.comfy_patched_weights = True
@@ -825,36 +760,17 @@ class ModelPatcher:
             for x in load_completely:
                 x[2].to(device_to)
 
-            # handle flipflop
-            if len(load_flipflop) > 0:
-                start_time = time.perf_counter()
-                load_flipflop.sort(reverse=True)
-                for x in load_flipflop:
-                    n = x[1]
-                    m = x[2]
-                    params = x[3]
-                    if hasattr(m, "comfy_patched_weights"):
-                        if m.comfy_patched_weights == True:
-                            continue
-                    for param in params:
-                        self.patch_weight_to_device("{}.{}".format(n, param), device_to=device_to, device_final=self.offload_device)
+            for x in offloaded:
+                n = x[1]
+                params = x[3]
+                for param in params:
+                    self.pin_weight_to_device("{}.{}".format(n, param))
 
-                    logging.debug("lowvram: loaded module for flipflop {} {}".format(n, m))
-                end_time = time.perf_counter()
-                logging.info(f"flipflop load time: {end_time - start_time:.2f} seconds")
-                start_time = time.perf_counter()
-                mem_counter += self.init_flipflop_block_copies()
-                end_time = time.perf_counter()
-                logging.info(f"flipflop block init time: {end_time - start_time:.2f} seconds")
-
-            if lowvram_counter > 0 or flipflop_counter > 0:
-                if flipflop_counter > 0:
-                    logging.info(f"loaded partially; {lowvram_model_memory / (1024 * 1024):.2f} MB usable, {mem_counter / (1024 * 1024):.2f} MB loaded, {flipflop_mem_counter / (1024 * 1024):.2f} MB flipflop, {lowvram_mem_counter / (1024 * 1024):.2f} MB offloaded, lowvram patches: {patch_counter}")
-                else:
-                    logging.info(f"loaded partially; {lowvram_model_memory / (1024 * 1024):.2f} MB usable, {mem_counter / (1024 * 1024):.2f} MB loaded, {lowvram_mem_counter / (1024 * 1024):.2f} MB offloaded, lowvram patches: {patch_counter}")
+            if lowvram_counter > 0:
+                logging.info("loaded partially; {:.2f} MB usable, {:.2f} MB loaded, {:.2f} MB offloaded, lowvram patches: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), patch_counter))
                 self.model.model_lowvram = True
             else:
-                logging.info(f"loaded completely; {lowvram_model_memory / (1024 * 1024):.2f} MB usable, {mem_counter / (1024 * 1024):.2f} MB loaded, full load: {full_load}")
+                logging.info("loaded completely; {:.2f} MB usable, {:.2f} MB loaded, full load: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), full_load))
                 self.model.model_lowvram = False
                 if full_load:
                     self.model.to(device_to)
@@ -891,7 +807,7 @@ class ModelPatcher:
         self.eject_model()
         if unpatch_weights:
             self.unpatch_hooks()
-            self.clean_flipflop()
+            self.unpin_all_weights()
             if self.model.model_lowvram:
                 for m in self.model.modules():
                     move_weight_functions(m, device_to)
@@ -931,9 +847,8 @@ class ModelPatcher:
         with self.use_ejected():
             hooks_unpatched = False
             memory_freed = 0
-            memory_freed += self.clean_flipflop()
             patch_counter = 0
-            unload_list = self._load_list(get_existing_flipflop=True)
+            unload_list = self._load_list()
             unload_list.sort()
             for unload in unload_list:
                 if memory_to_free < memory_freed:
@@ -942,10 +857,7 @@ class ModelPatcher:
                 n = unload[1]
                 m = unload[2]
                 params = unload[3]
-                flipflop: bool = unload[4]
 
-                if flipflop:
-                    continue
                 lowvram_possible = hasattr(m, "comfy_cast_weights")
                 if hasattr(m, "comfy_patched_weights") and m.comfy_patched_weights == True:
                     move_weight = True
@@ -991,6 +903,9 @@ class ModelPatcher:
                         memory_freed += module_mem
                         logging.debug("freed {}".format(n))
 
+                        for param in params:
+                            self.pin_weight_to_device("{}.{}".format(n, param))
+
             self.model.model_lowvram = True
             self.model.lowvram_patch_counter += patch_counter
             self.model.model_loaded_weight_memory -= memory_freed
@@ -1393,5 +1308,6 @@ class ModelPatcher:
         self.clear_cached_hook_weights()
 
     def __del__(self):
+        self.unpin_all_weights()
         self.detach(unpatch_all=False)
 

comfy/ops.py

@@ -35,7 +35,7 @@ def scaled_dot_product_attention(q, k, v, *args, **kwargs):
 
 
 try:
-    if torch.cuda.is_available():
+    if torch.cuda.is_available() and comfy.model_management.WINDOWS:
         from torch.nn.attention import SDPBackend, sdpa_kernel
         import inspect
         if "set_priority" in inspect.signature(sdpa_kernel).parameters:
@@ -70,8 +70,11 @@ cast_to = comfy.model_management.cast_to #TODO: remove once no more references
 def cast_to_input(weight, input, non_blocking=False, copy=True):
     return comfy.model_management.cast_to(weight, input.dtype, input.device, non_blocking=non_blocking, copy=copy)
 
-@torch.compiler.disable()
-def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None):
+
+def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, offloadable=False):
+    # NOTE: offloadable=False is a a legacy and if you are a custom node author reading this please pass
+    # offloadable=True and call uncast_bias_weight() after your last usage of the weight/bias. This
+    # will add async-offload support to your cast and improve performance.
     if input is not None:
         if dtype is None:
             dtype = input.dtype
@@ -80,32 +83,58 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None):
         if device is None:
             device = input.device
 
-    offload_stream = comfy.model_management.get_offload_stream(device)
+    if offloadable and (device != s.weight.device or
+                        (s.bias is not None and device != s.bias.device)):
+        offload_stream = comfy.model_management.get_offload_stream(device)
+    else:
+        offload_stream = None
+
     if offload_stream is not None:
         wf_context = offload_stream
     else:
         wf_context = contextlib.nullcontext()
 
-    bias = None
     non_blocking = comfy.model_management.device_supports_non_blocking(device)
-    if s.bias is not None:
-        has_function = len(s.bias_function) > 0
-        bias = comfy.model_management.cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=has_function, stream=offload_stream)
 
-        if has_function:
+    weight_has_function = len(s.weight_function) > 0
+    bias_has_function = len(s.bias_function) > 0
+
+    weight = comfy.model_management.cast_to(s.weight, None, device, non_blocking=non_blocking, copy=weight_has_function, stream=offload_stream)
+
+    bias = None
+    if s.bias is not None:
+        bias = comfy.model_management.cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=bias_has_function, stream=offload_stream)
+
+        if bias_has_function:
             with wf_context:
                 for f in s.bias_function:
                     bias = f(bias)
 
-    has_function = len(s.weight_function) > 0
-    weight = comfy.model_management.cast_to(s.weight, dtype, device, non_blocking=non_blocking, copy=has_function, stream=offload_stream)
-    if has_function:
+    weight = weight.to(dtype=dtype)
+    if weight_has_function:
         with wf_context:
             for f in s.weight_function:
                 weight = f(weight)
 
     comfy.model_management.sync_stream(device, offload_stream)
-    return weight, bias
+    if offloadable:
+        return weight, bias, offload_stream
+    else:
+        #Legacy function signature
+        return weight, bias
+
+
+def uncast_bias_weight(s, weight, bias, offload_stream):
+    if offload_stream is None:
+        return
+    if weight is not None:
+        device = weight.device
+    else:
+        if bias is None:
+            return
+        device = bias.device
+    offload_stream.wait_stream(comfy.model_management.current_stream(device))
+
 
 class CastWeightBiasOp:
     comfy_cast_weights = False
@@ -118,8 +147,10 @@ class disable_weight_init:
             return None
 
         def forward_comfy_cast_weights(self, input):
-            weight, bias = cast_bias_weight(self, input)
-            return torch.nn.functional.linear(input, weight, bias)
+            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+            x = torch.nn.functional.linear(input, weight, bias)
+            uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
         def forward(self, *args, **kwargs):
             run_every_op()
@@ -133,8 +164,10 @@ class disable_weight_init:
             return None
 
         def forward_comfy_cast_weights(self, input):
-            weight, bias = cast_bias_weight(self, input)
-            return self._conv_forward(input, weight, bias)
+            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+            x = self._conv_forward(input, weight, bias)
+            uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
         def forward(self, *args, **kwargs):
             run_every_op()
@@ -148,8 +181,10 @@ class disable_weight_init:
             return None
 
         def forward_comfy_cast_weights(self, input):
-            weight, bias = cast_bias_weight(self, input)
-            return self._conv_forward(input, weight, bias)
+            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+            x = self._conv_forward(input, weight, bias)
+            uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
         def forward(self, *args, **kwargs):
             run_every_op()
@@ -172,8 +207,10 @@ class disable_weight_init:
                 return super()._conv_forward(input, weight, bias, *args, **kwargs)
 
         def forward_comfy_cast_weights(self, input):
-            weight, bias = cast_bias_weight(self, input)
-            return self._conv_forward(input, weight, bias)
+            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+            x = self._conv_forward(input, weight, bias)
+            uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
         def forward(self, *args, **kwargs):
             run_every_op()
@@ -187,8 +224,10 @@ class disable_weight_init:
             return None
 
         def forward_comfy_cast_weights(self, input):
-            weight, bias = cast_bias_weight(self, input)
-            return torch.nn.functional.group_norm(input, self.num_groups, weight, bias, self.eps)
+            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+            x = torch.nn.functional.group_norm(input, self.num_groups, weight, bias, self.eps)
+            uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
         def forward(self, *args, **kwargs):
             run_every_op()
@@ -203,11 +242,14 @@ class disable_weight_init:
 
         def forward_comfy_cast_weights(self, input):
             if self.weight is not None:
-                weight, bias = cast_bias_weight(self, input)
+                weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
             else:
                 weight = None
                 bias = None
-            return torch.nn.functional.layer_norm(input, self.normalized_shape, weight, bias, self.eps)
+                offload_stream = None
+            x = torch.nn.functional.layer_norm(input, self.normalized_shape, weight, bias, self.eps)
+            uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
         def forward(self, *args, **kwargs):
             run_every_op()
@@ -223,11 +265,15 @@ class disable_weight_init:
 
         def forward_comfy_cast_weights(self, input):
             if self.weight is not None:
-                weight, bias = cast_bias_weight(self, input)
+                weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
             else:
                 weight = None
-            return comfy.rmsnorm.rms_norm(input, weight, self.eps)  # TODO: switch to commented out line when old torch is deprecated
-            # return torch.nn.functional.rms_norm(input, self.normalized_shape, weight, self.eps)
+                bias = None
+                offload_stream = None
+            x = comfy.rmsnorm.rms_norm(input, weight, self.eps)  # TODO: switch to commented out line when old torch is deprecated
+            # x = torch.nn.functional.rms_norm(input, self.normalized_shape, weight, self.eps)
+            uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
         def forward(self, *args, **kwargs):
             run_every_op()
@@ -246,10 +292,12 @@ class disable_weight_init:
                 input, output_size, self.stride, self.padding, self.kernel_size,
                 num_spatial_dims, self.dilation)
 
-            weight, bias = cast_bias_weight(self, input)
-            return torch.nn.functional.conv_transpose2d(
+            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+            x = torch.nn.functional.conv_transpose2d(
                 input, weight, bias, self.stride, self.padding,
                 output_padding, self.groups, self.dilation)
+            uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
         def forward(self, *args, **kwargs):
             run_every_op()
@@ -268,10 +316,12 @@ class disable_weight_init:
                 input, output_size, self.stride, self.padding, self.kernel_size,
                 num_spatial_dims, self.dilation)
 
-            weight, bias = cast_bias_weight(self, input)
-            return torch.nn.functional.conv_transpose1d(
+            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+            x = torch.nn.functional.conv_transpose1d(
                 input, weight, bias, self.stride, self.padding,
                 output_padding, self.groups, self.dilation)
+            uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
         def forward(self, *args, **kwargs):
             run_every_op()
@@ -289,8 +339,11 @@ class disable_weight_init:
             output_dtype = out_dtype
             if self.weight.dtype == torch.float16 or self.weight.dtype == torch.bfloat16:
                 out_dtype = None
-            weight, bias = cast_bias_weight(self, device=input.device, dtype=out_dtype)
-            return torch.nn.functional.embedding(input, weight, self.padding_idx, self.max_norm, self.norm_type, self.scale_grad_by_freq, self.sparse).to(dtype=output_dtype)
+            weight, bias, offload_stream = cast_bias_weight(self, device=input.device, dtype=out_dtype, offloadable=True)
+            x = torch.nn.functional.embedding(input, weight, self.padding_idx, self.max_norm, self.norm_type, self.scale_grad_by_freq, self.sparse).to(dtype=output_dtype)
+            uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
+
 
         def forward(self, *args, **kwargs):
             run_every_op()
@@ -352,16 +405,10 @@ def fp8_linear(self, input):
     if dtype not in [torch.float8_e4m3fn]:
         return None
 
-    tensor_2d = False
-    if len(input.shape) == 2:
-        tensor_2d = True
-        input = input.unsqueeze(1)
-
-    input_shape = input.shape
     input_dtype = input.dtype
 
-    if len(input.shape) == 3:
-        w, bias = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input_dtype)
+    if input.ndim == 3 or input.ndim == 2:
+        w, bias, offload_stream = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input_dtype, offloadable=True)
 
         scale_weight = self.scale_weight
         scale_input = self.scale_input
@@ -372,19 +419,21 @@ def fp8_linear(self, input):
 
         if scale_input is None:
             scale_input = torch.ones((), device=input.device, dtype=torch.float32)
+            input = torch.clamp(input, min=-448, max=448, out=input)
+            layout_params_weight = {'scale': scale_input, 'orig_dtype': input_dtype}
+            quantized_input = QuantizedTensor(input.to(dtype).contiguous(), "TensorCoreFP8Layout", layout_params_weight)
         else:
             scale_input = scale_input.to(input.device)
+            quantized_input = QuantizedTensor.from_float(input, "TensorCoreFP8Layout", scale=scale_input, dtype=dtype)
 
         # Wrap weight in QuantizedTensor - this enables unified dispatch
         # Call F.linear - __torch_dispatch__ routes to fp8_linear handler in quant_ops.py!
         layout_params_weight = {'scale': scale_weight, 'orig_dtype': input_dtype}
-        quantized_weight = QuantizedTensor(w, TensorCoreFP8Layout, layout_params_weight)
-        quantized_input = QuantizedTensor.from_float(input.reshape(-1, input_shape[2]), TensorCoreFP8Layout, scale=scale_input, dtype=dtype)
+        quantized_weight = QuantizedTensor(w, "TensorCoreFP8Layout", layout_params_weight)
         o = torch.nn.functional.linear(quantized_input, quantized_weight, bias)
 
-        if tensor_2d:
-            return o.reshape(input_shape[0], -1)
-        return o.reshape((-1, input_shape[1], self.weight.shape[0]))
+        uncast_bias_weight(self, w, bias, offload_stream)
+        return o
 
     return None
 
@@ -404,8 +453,10 @@ class fp8_ops(manual_cast):
                 except Exception as e:
                     logging.info("Exception during fp8 op: {}".format(e))
 
-            weight, bias = cast_bias_weight(self, input)
-            return torch.nn.functional.linear(input, weight, bias)
+            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+            x = torch.nn.functional.linear(input, weight, bias)
+            uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
 def scaled_fp8_ops(fp8_matrix_mult=False, scale_input=False, override_dtype=None):
     logging.info("Using scaled fp8: fp8 matrix mult: {}, scale input: {}".format(fp8_matrix_mult, scale_input))
@@ -433,12 +484,14 @@ def scaled_fp8_ops(fp8_matrix_mult=False, scale_input=False, override_dtype=None
                     if out is not None:
                         return out
 
-                weight, bias = cast_bias_weight(self, input)
+                weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
 
                 if weight.numel() < input.numel(): #TODO: optimize
-                    return torch.nn.functional.linear(input, weight * self.scale_weight.to(device=weight.device, dtype=weight.dtype), bias)
+                    x = torch.nn.functional.linear(input, weight * self.scale_weight.to(device=weight.device, dtype=weight.dtype), bias)
                 else:
-                    return torch.nn.functional.linear(input * self.scale_weight.to(device=weight.device, dtype=weight.dtype), weight, bias)
+                    x = torch.nn.functional.linear(input * self.scale_weight.to(device=weight.device, dtype=weight.dtype), weight, bias)
+                uncast_bias_weight(self, weight, bias, offload_stream)
+                return x
 
             def convert_weight(self, weight, inplace=False, **kwargs):
                 if inplace:
@@ -481,12 +534,12 @@ if CUBLAS_IS_AVAILABLE:
 # ==============================================================================
 # Mixed Precision Operations
 # ==============================================================================
-from .quant_ops import QuantizedTensor, TensorCoreFP8Layout
+from .quant_ops import QuantizedTensor
 
 QUANT_FORMAT_MIXINS = {
     "float8_e4m3fn": {
         "dtype": torch.float8_e4m3fn,
-        "layout_type": TensorCoreFP8Layout,
+        "layout_type": "TensorCoreFP8Layout",
         "parameters": {
             "weight_scale": torch.nn.Parameter(torch.zeros((), dtype=torch.float32), requires_grad=False),
             "input_scale": torch.nn.Parameter(torch.zeros((), dtype=torch.float32), requires_grad=False),
@@ -577,8 +630,10 @@ class MixedPrecisionOps(disable_weight_init):
             return torch.nn.functional.linear(input, weight, bias)
 
         def forward_comfy_cast_weights(self, input):
-            weight, bias = cast_bias_weight(self, input)
-            return self._forward(input, weight, bias)
+            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+            x = self._forward(input, weight, bias)
+            uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
         def forward(self, input, *args, **kwargs):
             run_every_op()

comfy/quant_ops.py

@@ -123,15 +123,15 @@ class QuantizedTensor(torch.Tensor):
             layout_type: Layout class (subclass of QuantizedLayout)
             layout_params: Dict with layout-specific parameters
         """
-        return torch.Tensor._make_subclass(cls, qdata, require_grad=False)
+        return torch.Tensor._make_wrapper_subclass(cls, qdata.shape, device=qdata.device, dtype=qdata.dtype, requires_grad=False)
 
     def __init__(self, qdata, layout_type, layout_params):
-        self._qdata = qdata.contiguous()
+        self._qdata = qdata
         self._layout_type = layout_type
         self._layout_params = layout_params
 
     def __repr__(self):
-        layout_name = self._layout_type.__name__
+        layout_name = self._layout_type
         param_str = ", ".join(f"{k}={v}" for k, v in list(self._layout_params.items())[:2])
         return f"QuantizedTensor(shape={self.shape}, layout={layout_name}, {param_str})"
 
@@ -179,15 +179,15 @@ class QuantizedTensor(torch.Tensor):
             attr_name = f"_layout_param_{key}"
             layout_params[key] = inner_tensors[attr_name]
 
-        return QuantizedTensor(inner_tensors["_q_data"], layout_type, layout_params)
+        return QuantizedTensor(inner_tensors["_qdata"], layout_type, layout_params)
 
     @classmethod
     def from_float(cls, tensor, layout_type, **quantize_kwargs) -> 'QuantizedTensor':
-        qdata, layout_params = layout_type.quantize(tensor, **quantize_kwargs)
+        qdata, layout_params = LAYOUTS[layout_type].quantize(tensor, **quantize_kwargs)
         return cls(qdata, layout_type, layout_params)
 
     def dequantize(self) -> torch.Tensor:
-        return self._layout_type.dequantize(self._qdata, **self._layout_params)
+        return LAYOUTS[self._layout_type].dequantize(self._qdata, **self._layout_params)
 
     @classmethod
     def __torch_dispatch__(cls, func, types, args=(), kwargs=None):
@@ -357,9 +357,10 @@ class TensorCoreFP8Layout(QuantizedLayout):
             scale = torch.tensor(scale)
         scale = scale.to(device=tensor.device, dtype=torch.float32)
 
-        lp_amax = torch.finfo(dtype).max
-        tensor_scaled = tensor.float() / scale
-        torch.clamp(tensor_scaled, min=-lp_amax, max=lp_amax, out=tensor_scaled)
+        tensor_scaled = tensor * (1.0 / scale).to(tensor.dtype)
+        # TODO: uncomment this if it's actually needed because the clamp has a small performance penality'
+        # lp_amax = torch.finfo(dtype).max
+        # torch.clamp(tensor_scaled, min=-lp_amax, max=lp_amax, out=tensor_scaled)
         qdata = tensor_scaled.to(dtype, memory_format=torch.contiguous_format)
 
         layout_params = {
@@ -378,7 +379,12 @@ class TensorCoreFP8Layout(QuantizedLayout):
         return qtensor._qdata, qtensor._layout_params['scale']
 
 
-@register_layout_op(torch.ops.aten.linear.default, TensorCoreFP8Layout)
+LAYOUTS = {
+    "TensorCoreFP8Layout": TensorCoreFP8Layout,
+}
+
+
+@register_layout_op(torch.ops.aten.linear.default, "TensorCoreFP8Layout")
 def fp8_linear(func, args, kwargs):
     input_tensor = args[0]
     weight = args[1]
@@ -405,13 +411,17 @@ def fp8_linear(func, args, kwargs):
 
         try:
             output = torch._scaled_mm(
-                plain_input.reshape(-1, input_shape[2]),
+                plain_input.reshape(-1, input_shape[2]).contiguous(),
                 weight_t,
                 bias=bias,
                 scale_a=scale_a,
                 scale_b=scale_b,
                 out_dtype=out_dtype,
             )
+
+            if isinstance(output, tuple):  # TODO: remove when we drop support for torch 2.4
+                output = output[0]
+
             if not tensor_2d:
                 output = output.reshape((-1, input_shape[1], weight.shape[0]))
 
@@ -421,7 +431,7 @@ def fp8_linear(func, args, kwargs):
                     'scale': output_scale,
                     'orig_dtype': input_tensor._layout_params['orig_dtype']
                 }
-                return QuantizedTensor(output, TensorCoreFP8Layout, output_params)
+                return QuantizedTensor(output, "TensorCoreFP8Layout", output_params)
             else:
                 return output
 
@@ -435,3 +445,68 @@ def fp8_linear(func, args, kwargs):
         input_tensor = input_tensor.dequantize()
 
     return torch.nn.functional.linear(input_tensor, weight, bias)
+
+def fp8_mm_(input_tensor, weight, bias=None, out_dtype=None):
+    if out_dtype is None:
+        out_dtype = input_tensor._layout_params['orig_dtype']
+
+    plain_input, scale_a = TensorCoreFP8Layout.get_plain_tensors(input_tensor)
+    plain_weight, scale_b = TensorCoreFP8Layout.get_plain_tensors(weight)
+
+    output = torch._scaled_mm(
+        plain_input.contiguous(),
+        plain_weight,
+        bias=bias,
+        scale_a=scale_a,
+        scale_b=scale_b,
+        out_dtype=out_dtype,
+    )
+
+    if isinstance(output, tuple):  # TODO: remove when we drop support for torch 2.4
+        output = output[0]
+    return output
+
+@register_layout_op(torch.ops.aten.addmm.default, "TensorCoreFP8Layout")
+def fp8_addmm(func, args, kwargs):
+    input_tensor = args[1]
+    weight = args[2]
+    bias = args[0]
+
+    if isinstance(input_tensor, QuantizedTensor) and isinstance(weight, QuantizedTensor):
+        return fp8_mm_(input_tensor, weight, bias=bias, out_dtype=kwargs.get("out_dtype", None))
+
+    a = list(args)
+    if isinstance(args[0], QuantizedTensor):
+        a[0] = args[0].dequantize()
+    if isinstance(args[1], QuantizedTensor):
+        a[1] = args[1].dequantize()
+    if isinstance(args[2], QuantizedTensor):
+        a[2] = args[2].dequantize()
+
+    return func(*a, **kwargs)
+
+@register_layout_op(torch.ops.aten.mm.default, "TensorCoreFP8Layout")
+def fp8_mm(func, args, kwargs):
+    input_tensor = args[0]
+    weight = args[1]
+
+    if isinstance(input_tensor, QuantizedTensor) and isinstance(weight, QuantizedTensor):
+        return fp8_mm_(input_tensor, weight, bias=None, out_dtype=kwargs.get("out_dtype", None))
+
+    a = list(args)
+    if isinstance(args[0], QuantizedTensor):
+        a[0] = args[0].dequantize()
+    if isinstance(args[1], QuantizedTensor):
+        a[1] = args[1].dequantize()
+    return func(*a, **kwargs)
+
+@register_layout_op(torch.ops.aten.view.default, "TensorCoreFP8Layout")
+@register_layout_op(torch.ops.aten.t.default, "TensorCoreFP8Layout")
+def fp8_func(func, args, kwargs):
+    input_tensor = args[0]
+    if isinstance(input_tensor, QuantizedTensor):
+        plain_input, scale_a = TensorCoreFP8Layout.get_plain_tensors(input_tensor)
+        ar = list(args)
+        ar[0] = plain_input
+        return QuantizedTensor(func(*ar, **kwargs), "TensorCoreFP8Layout", input_tensor._layout_params)
+    return func(*args, **kwargs)

comfy/sd.py

@@ -143,6 +143,9 @@ class CLIP:
         n.apply_hooks_to_conds = self.apply_hooks_to_conds
         return n
 
+    def get_ram_usage(self):
+        return self.patcher.get_ram_usage()
+
     def add_patches(self, patches, strength_patch=1.0, strength_model=1.0):
         return self.patcher.add_patches(patches, strength_patch, strength_model)
 
@@ -293,6 +296,7 @@ class VAE:
         self.working_dtypes = [torch.bfloat16, torch.float32]
         self.disable_offload = False
         self.not_video = False
+        self.size = None
 
         self.downscale_index_formula = None
         self.upscale_index_formula = None
@@ -595,6 +599,16 @@ class VAE:
 
         self.patcher = comfy.model_patcher.ModelPatcher(self.first_stage_model, load_device=self.device, offload_device=offload_device)
         logging.info("VAE load device: {}, offload device: {}, dtype: {}".format(self.device, offload_device, self.vae_dtype))
+        self.model_size()
+
+    def model_size(self):
+        if self.size is not None:
+            return self.size
+        self.size = comfy.model_management.module_size(self.first_stage_model)
+        return self.size
+
+    def get_ram_usage(self):
+        return self.model_size()
 
     def throw_exception_if_invalid(self):
         if self.first_stage_model is None:
@@ -1330,7 +1344,7 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None):
     else:
         unet_dtype = dtype
 
-    if hasattr(model_config, "layer_quant_config"):
+    if model_config.layer_quant_config is not None:
         manual_cast_dtype = model_management.unet_manual_cast(None, load_device, model_config.supported_inference_dtypes)
     else:
         manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device, model_config.supported_inference_dtypes)

comfy_api_nodes/apinode_utils.py

@@ -1,261 +0,0 @@
-from __future__ import annotations
-import aiohttp
-import mimetypes
-from typing import Optional, Union
-from comfy.utils import common_upscale
-from comfy_api_nodes.apis.client import (
-    ApiClient,
-    ApiEndpoint,
-    HttpMethod,
-    SynchronousOperation,
-    UploadRequest,
-    UploadResponse,
-)
-from server import PromptServer
-from comfy.cli_args import args
-
-import numpy as np
-from PIL import Image
-import torch
-import math
-import base64
-from .util import tensor_to_bytesio, bytesio_to_image_tensor
-from io import BytesIO
-
-
-async def validate_and_cast_response(
-    response, timeout: int = None, node_id: Union[str, None] = None
-) -> torch.Tensor:
-    """Validates and casts a response to a torch.Tensor.
-
-    Args:
-        response: The response to validate and cast.
-        timeout: Request timeout in seconds. Defaults to None (no timeout).
-
-    Returns:
-        A torch.Tensor representing the image (1, H, W, C).
-
-    Raises:
-        ValueError: If the response is not valid.
-    """
-    # validate raw JSON response
-    data = response.data
-    if not data or len(data) == 0:
-        raise ValueError("No images returned from API endpoint")
-
-    # Initialize list to store image tensors
-    image_tensors: list[torch.Tensor] = []
-
-    # Process each image in the data array
-    async with aiohttp.ClientSession(timeout=aiohttp.ClientTimeout(total=timeout)) as session:
-        for img_data in data:
-            img_bytes: bytes
-            if img_data.b64_json:
-                img_bytes = base64.b64decode(img_data.b64_json)
-            elif img_data.url:
-                if node_id:
-                    PromptServer.instance.send_progress_text(f"Result URL: {img_data.url}", node_id)
-                async with session.get(img_data.url) as resp:
-                    if resp.status != 200:
-                        raise ValueError("Failed to download generated image")
-                    img_bytes = await resp.read()
-            else:
-                raise ValueError("Invalid image payload – neither URL nor base64 data present.")
-
-            pil_img = Image.open(BytesIO(img_bytes)).convert("RGBA")
-            arr = np.asarray(pil_img).astype(np.float32) / 255.0
-            image_tensors.append(torch.from_numpy(arr))
-
-    return torch.stack(image_tensors, dim=0)
-
-
-def validate_aspect_ratio(
-    aspect_ratio: str,
-    minimum_ratio: float,
-    maximum_ratio: float,
-    minimum_ratio_str: str,
-    maximum_ratio_str: str,
-) -> float:
-    """Validates and casts an aspect ratio string to a float.
-
-    Args:
-        aspect_ratio: The aspect ratio string to validate.
-        minimum_ratio: The minimum aspect ratio.
-        maximum_ratio: The maximum aspect ratio.
-        minimum_ratio_str: The minimum aspect ratio string.
-        maximum_ratio_str: The maximum aspect ratio string.
-
-    Returns:
-        The validated and cast aspect ratio.
-
-    Raises:
-        Exception: If the aspect ratio is not valid.
-    """
-    # get ratio values
-    numbers = aspect_ratio.split(":")
-    if len(numbers) != 2:
-        raise TypeError(
-            f"Aspect ratio must be in the format X:Y, such as 16:9, but was {aspect_ratio}."
-        )
-    try:
-        numerator = int(numbers[0])
-        denominator = int(numbers[1])
-    except ValueError as exc:
-        raise TypeError(
-            f"Aspect ratio must contain numbers separated by ':', such as 16:9, but was {aspect_ratio}."
-        ) from exc
-    calculated_ratio = numerator / denominator
-    # if not close to minimum and maximum, check bounds
-    if not math.isclose(calculated_ratio, minimum_ratio) or not math.isclose(
-        calculated_ratio, maximum_ratio
-    ):
-        if calculated_ratio < minimum_ratio:
-            raise TypeError(
-                f"Aspect ratio cannot reduce to any less than {minimum_ratio_str} ({minimum_ratio}), but was {aspect_ratio} ({calculated_ratio})."
-            )
-        if calculated_ratio > maximum_ratio:
-            raise TypeError(
-                f"Aspect ratio cannot reduce to any greater than {maximum_ratio_str} ({maximum_ratio}), but was {aspect_ratio} ({calculated_ratio})."
-            )
-    return aspect_ratio
-
-
-async def download_url_to_bytesio(
-    url: str, timeout: int = None, auth_kwargs: Optional[dict[str, str]] = None
-) -> BytesIO:
-    """Downloads content from a URL using requests and returns it as BytesIO.
-
-    Args:
-        url: The URL to download.
-        timeout: Request timeout in seconds. Defaults to None (no timeout).
-
-    Returns:
-        BytesIO object containing the downloaded content.
-    """
-    headers = {}
-    if url.startswith("/proxy/"):
-        url = str(args.comfy_api_base).rstrip("/") + url
-        auth_token = auth_kwargs.get("auth_token")
-        comfy_api_key = auth_kwargs.get("comfy_api_key")
-        if auth_token:
-            headers["Authorization"] = f"Bearer {auth_token}"
-        elif comfy_api_key:
-            headers["X-API-KEY"] = comfy_api_key
-    timeout_cfg = aiohttp.ClientTimeout(total=timeout) if timeout else None
-    async with aiohttp.ClientSession(timeout=timeout_cfg) as session:
-        async with session.get(url, headers=headers) as resp:
-            resp.raise_for_status()  # Raises HTTPError for bad responses (4XX or 5XX)
-            return BytesIO(await resp.read())
-
-
-def process_image_response(response_content: bytes | str) -> torch.Tensor:
-    """Uses content from a Response object and converts it to a torch.Tensor"""
-    return bytesio_to_image_tensor(BytesIO(response_content))
-
-
-def text_filepath_to_base64_string(filepath: str) -> str:
-    """Converts a text file to a base64 string."""
-    with open(filepath, "rb") as f:
-        file_content = f.read()
-    return base64.b64encode(file_content).decode("utf-8")
-
-
-def text_filepath_to_data_uri(filepath: str) -> str:
-    """Converts a text file to a data URI."""
-    base64_string = text_filepath_to_base64_string(filepath)
-    mime_type, _ = mimetypes.guess_type(filepath)
-    if mime_type is None:
-        mime_type = "application/octet-stream"
-    return f"data:{mime_type};base64,{base64_string}"
-
-
-async def upload_file_to_comfyapi(
-    file_bytes_io: BytesIO,
-    filename: str,
-    upload_mime_type: Optional[str],
-    auth_kwargs: Optional[dict[str, str]] = None,
-) -> str:
-    """
-    Uploads a single file to ComfyUI API and returns its download URL.
-
-    Args:
-        file_bytes_io: BytesIO object containing the file data.
-        filename: The filename of the file.
-        upload_mime_type: MIME type of the file.
-        auth_kwargs: Optional authentication token(s).
-
-    Returns:
-        The download URL for the uploaded file.
-    """
-    if upload_mime_type is None:
-        request_object = UploadRequest(file_name=filename)
-    else:
-        request_object = UploadRequest(file_name=filename, content_type=upload_mime_type)
-    operation = SynchronousOperation(
-        endpoint=ApiEndpoint(
-            path="/customers/storage",
-            method=HttpMethod.POST,
-            request_model=UploadRequest,
-            response_model=UploadResponse,
-        ),
-        request=request_object,
-        auth_kwargs=auth_kwargs,
-    )
-
-    response: UploadResponse = await operation.execute()
-    await ApiClient.upload_file(response.upload_url, file_bytes_io, content_type=upload_mime_type)
-    return response.download_url
-
-
-async def upload_images_to_comfyapi(
-    image: torch.Tensor,
-    max_images=8,
-    auth_kwargs: Optional[dict[str, str]] = None,
-    mime_type: Optional[str] = None,
-) -> list[str]:
-    """
-    Uploads images to ComfyUI API and returns download URLs.
-    To upload multiple images, stack them in the batch dimension first.
-
-    Args:
-        image: Input torch.Tensor image.
-        max_images: Maximum number of images to upload.
-        auth_kwargs: Optional authentication token(s).
-        mime_type: Optional MIME type for the image.
-    """
-    # if batch, try to upload each file if max_images is greater than 0
-    download_urls: list[str] = []
-    is_batch = len(image.shape) > 3
-    batch_len = image.shape[0] if is_batch else 1
-
-    for idx in range(min(batch_len, max_images)):
-        tensor = image[idx] if is_batch else image
-        img_io = tensor_to_bytesio(tensor, mime_type=mime_type)
-        url = await upload_file_to_comfyapi(img_io, img_io.name, mime_type, auth_kwargs)
-        download_urls.append(url)
-    return download_urls
-
-
-def resize_mask_to_image(
-    mask: torch.Tensor,
-    image: torch.Tensor,
-    upscale_method="nearest-exact",
-    crop="disabled",
-    allow_gradient=True,
-    add_channel_dim=False,
-):
-    """
-    Resize mask to be the same dimensions as an image, while maintaining proper format for API calls.
-    """
-    _, H, W, _ = image.shape
-    mask = mask.unsqueeze(-1)
-    mask = mask.movedim(-1, 1)
-    mask = common_upscale(
-        mask, width=W, height=H, upscale_method=upscale_method, crop=crop
-    )
-    mask = mask.movedim(1, -1)
-    if not add_channel_dim:
-        mask = mask.squeeze(-1)
-    if not allow_gradient:
-        mask = (mask > 0.5).float()
-    return mask

comfy_api_nodes/apis/minimax_api.py

@@ -0,0 +1,120 @@
+from enum import Enum
+from typing import Optional
+
+from pydantic import BaseModel, Field
+
+
+class MinimaxBaseResponse(BaseModel):
+    status_code: int = Field(
+        ...,
+        description='Status code. 0 indicates success, other values indicate errors.',
+    )
+    status_msg: str = Field(
+        ..., description='Specific error details or success message.'
+    )
+
+
+class File(BaseModel):
+    bytes: Optional[int] = Field(None, description='File size in bytes')
+    created_at: Optional[int] = Field(
+        None, description='Unix timestamp when the file was created, in seconds'
+    )
+    download_url: Optional[str] = Field(
+        None, description='The URL to download the video'
+    )
+    backup_download_url: Optional[str] = Field(
+        None, description='The backup URL to download the video'
+    )
+
+    file_id: Optional[int] = Field(None, description='Unique identifier for the file')
+    filename: Optional[str] = Field(None, description='The name of the file')
+    purpose: Optional[str] = Field(None, description='The purpose of using the file')
+
+
+class MinimaxFileRetrieveResponse(BaseModel):
+    base_resp: MinimaxBaseResponse
+    file: File
+
+
+class MiniMaxModel(str, Enum):
+    T2V_01_Director = 'T2V-01-Director'
+    I2V_01_Director = 'I2V-01-Director'
+    S2V_01 = 'S2V-01'
+    I2V_01 = 'I2V-01'
+    I2V_01_live = 'I2V-01-live'
+    T2V_01 = 'T2V-01'
+    Hailuo_02 = 'MiniMax-Hailuo-02'
+
+
+class Status6(str, Enum):
+    Queueing = 'Queueing'
+    Preparing = 'Preparing'
+    Processing = 'Processing'
+    Success = 'Success'
+    Fail = 'Fail'
+
+
+class MinimaxTaskResultResponse(BaseModel):
+    base_resp: MinimaxBaseResponse
+    file_id: Optional[str] = Field(
+        None,
+        description='After the task status changes to Success, this field returns the file ID corresponding to the generated video.',
+    )
+    status: Status6 = Field(
+        ...,
+        description="Task status: 'Queueing' (in queue), 'Preparing' (task is preparing), 'Processing' (generating), 'Success' (task completed successfully), or 'Fail' (task failed).",
+    )
+    task_id: str = Field(..., description='The task ID being queried.')
+
+
+class SubjectReferenceItem(BaseModel):
+    image: Optional[str] = Field(
+        None, description='URL or base64 encoding of the subject reference image.'
+    )
+    mask: Optional[str] = Field(
+        None,
+        description='URL or base64 encoding of the mask for the subject reference image.',
+    )
+
+
+class MinimaxVideoGenerationRequest(BaseModel):
+    callback_url: Optional[str] = Field(
+        None,
+        description='Optional. URL to receive real-time status updates about the video generation task.',
+    )
+    first_frame_image: Optional[str] = Field(
+        None,
+        description='URL or base64 encoding of the first frame image. Required when model is I2V-01, I2V-01-Director, or I2V-01-live.',
+    )
+    model: MiniMaxModel = Field(
+        ...,
+        description='Required. ID of model. Options: T2V-01-Director, I2V-01-Director, S2V-01, I2V-01, I2V-01-live, T2V-01',
+    )
+    prompt: Optional[str] = Field(
+        None,
+        description='Description of the video. Should be less than 2000 characters. Supports camera movement instructions in [brackets].',
+        max_length=2000,
+    )
+    prompt_optimizer: Optional[bool] = Field(
+        True,
+        description='If true (default), the model will automatically optimize the prompt. Set to false for more precise control.',
+    )
+    subject_reference: Optional[list[SubjectReferenceItem]] = Field(
+        None,
+        description='Only available when model is S2V-01. The model will generate a video based on the subject uploaded through this parameter.',
+    )
+    duration: Optional[int] = Field(
+        None,
+        description="The length of the output video in seconds."
+    )
+    resolution: Optional[str] = Field(
+        None,
+        description="The dimensions of the video display. 1080p corresponds to 1920 x 1080 pixels, 768p corresponds to 1366 x 768 pixels."
+    )
+
+
+class MinimaxVideoGenerationResponse(BaseModel):
+    base_resp: MinimaxBaseResponse
+    task_id: str = Field(
+        ..., description='The task ID for the asynchronous video generation task.'
+    )

comfy_api_nodes/nodes_bfl.py

@@ -5,10 +5,6 @@ import torch
 from typing_extensions import override
 
 from comfy_api.latest import IO, ComfyExtension
-from comfy_api_nodes.apinode_utils import (
-    resize_mask_to_image,
-    validate_aspect_ratio,
-)
 from comfy_api_nodes.apis.bfl_api import (
     BFLFluxExpandImageRequest,
     BFLFluxFillImageRequest,
@@ -23,8 +19,10 @@ from comfy_api_nodes.util import (
     ApiEndpoint,
     download_url_to_image_tensor,
     poll_op,
+    resize_mask_to_image,
     sync_op,
     tensor_to_base64_string,
+    validate_aspect_ratio_string,
     validate_string,
 )
 
@@ -43,11 +41,6 @@ class FluxProUltraImageNode(IO.ComfyNode):
     Generates images using Flux Pro 1.1 Ultra via api based on prompt and resolution.
     """
 
-    MINIMUM_RATIO = 1 / 4
-    MAXIMUM_RATIO = 4 / 1
-    MINIMUM_RATIO_STR = "1:4"
-    MAXIMUM_RATIO_STR = "4:1"
-
     @classmethod
     def define_schema(cls) -> IO.Schema:
         return IO.Schema(
@@ -112,16 +105,7 @@ class FluxProUltraImageNode(IO.ComfyNode):
 
     @classmethod
     def validate_inputs(cls, aspect_ratio: str):
-        try:
-            validate_aspect_ratio(
-                aspect_ratio,
-                minimum_ratio=cls.MINIMUM_RATIO,
-                maximum_ratio=cls.MAXIMUM_RATIO,
-                minimum_ratio_str=cls.MINIMUM_RATIO_STR,
-                maximum_ratio_str=cls.MAXIMUM_RATIO_STR,
-            )
-        except Exception as e:
-            return str(e)
+        validate_aspect_ratio_string(aspect_ratio, (1, 4), (4, 1))
         return True
 
     @classmethod
@@ -145,13 +129,7 @@ class FluxProUltraImageNode(IO.ComfyNode):
                 prompt=prompt,
                 prompt_upsampling=prompt_upsampling,
                 seed=seed,
-                aspect_ratio=validate_aspect_ratio(
-                    aspect_ratio,
-                    minimum_ratio=cls.MINIMUM_RATIO,
-                    maximum_ratio=cls.MAXIMUM_RATIO,
-                    minimum_ratio_str=cls.MINIMUM_RATIO_STR,
-                    maximum_ratio_str=cls.MAXIMUM_RATIO_STR,
-                ),
+                aspect_ratio=aspect_ratio,
                 raw=raw,
                 image_prompt=(image_prompt if image_prompt is None else tensor_to_base64_string(image_prompt)),
                 image_prompt_strength=(None if image_prompt is None else round(image_prompt_strength, 2)),
@@ -180,11 +158,6 @@ class FluxKontextProImageNode(IO.ComfyNode):
     Edits images using Flux.1 Kontext [pro] via api based on prompt and aspect ratio.
     """
 
-    MINIMUM_RATIO = 1 / 4
-    MAXIMUM_RATIO = 4 / 1
-    MINIMUM_RATIO_STR = "1:4"
-    MAXIMUM_RATIO_STR = "4:1"
-
     @classmethod
     def define_schema(cls) -> IO.Schema:
         return IO.Schema(
@@ -261,13 +234,7 @@ class FluxKontextProImageNode(IO.ComfyNode):
         seed=0,
         prompt_upsampling=False,
     ) -> IO.NodeOutput:
-        aspect_ratio = validate_aspect_ratio(
-            aspect_ratio,
-            minimum_ratio=cls.MINIMUM_RATIO,
-            maximum_ratio=cls.MAXIMUM_RATIO,
-            minimum_ratio_str=cls.MINIMUM_RATIO_STR,
-            maximum_ratio_str=cls.MAXIMUM_RATIO_STR,
-        )
+        validate_aspect_ratio_string(aspect_ratio, (1, 4), (4, 1))
         if input_image is None:
             validate_string(prompt, strip_whitespace=False)
         initial_response = await sync_op(

comfy_api_nodes/nodes_bytedance.py

@@ -17,7 +17,7 @@ from comfy_api_nodes.util import (
     poll_op,
     sync_op,
     upload_images_to_comfyapi,
-    validate_image_aspect_ratio_range,
+    validate_image_aspect_ratio,
     validate_image_dimensions,
     validate_string,
 )
@@ -403,7 +403,7 @@ class ByteDanceImageEditNode(IO.ComfyNode):
         validate_string(prompt, strip_whitespace=True, min_length=1)
         if get_number_of_images(image) != 1:
             raise ValueError("Exactly one input image is required.")
-        validate_image_aspect_ratio_range(image, (1, 3), (3, 1))
+        validate_image_aspect_ratio(image, (1, 3), (3, 1))
         source_url = (await upload_images_to_comfyapi(cls, image, max_images=1, mime_type="image/png"))[0]
         payload = Image2ImageTaskCreationRequest(
             model=model,
@@ -565,7 +565,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode):
         reference_images_urls = []
         if n_input_images:
             for i in image:
-                validate_image_aspect_ratio_range(i, (1, 3), (3, 1))
+                validate_image_aspect_ratio(i, (1, 3), (3, 1))
             reference_images_urls = await upload_images_to_comfyapi(
                 cls,
                 image,
@@ -798,7 +798,7 @@ class ByteDanceImageToVideoNode(IO.ComfyNode):
         validate_string(prompt, strip_whitespace=True, min_length=1)
         raise_if_text_params(prompt, ["resolution", "ratio", "duration", "seed", "camerafixed", "watermark"])
         validate_image_dimensions(image, min_width=300, min_height=300, max_width=6000, max_height=6000)
-        validate_image_aspect_ratio_range(image, (2, 5), (5, 2), strict=False)  # 0.4 to 2.5
+        validate_image_aspect_ratio(image, (2, 5), (5, 2), strict=False)  # 0.4 to 2.5
 
         image_url = (await upload_images_to_comfyapi(cls, image, max_images=1))[0]
         prompt = (
@@ -923,7 +923,7 @@ class ByteDanceFirstLastFrameNode(IO.ComfyNode):
         raise_if_text_params(prompt, ["resolution", "ratio", "duration", "seed", "camerafixed", "watermark"])
         for i in (first_frame, last_frame):
             validate_image_dimensions(i, min_width=300, min_height=300, max_width=6000, max_height=6000)
-            validate_image_aspect_ratio_range(i, (2, 5), (5, 2), strict=False)  # 0.4 to 2.5
+            validate_image_aspect_ratio(i, (2, 5), (5, 2), strict=False)  # 0.4 to 2.5
 
         download_urls = await upload_images_to_comfyapi(
             cls,
@@ -1045,7 +1045,7 @@ class ByteDanceImageReferenceNode(IO.ComfyNode):
         raise_if_text_params(prompt, ["resolution", "ratio", "duration", "seed", "watermark"])
         for image in images:
             validate_image_dimensions(image, min_width=300, min_height=300, max_width=6000, max_height=6000)
-            validate_image_aspect_ratio_range(image, (2, 5), (5, 2), strict=False)  # 0.4 to 2.5
+            validate_image_aspect_ratio(image, (2, 5), (5, 2), strict=False)  # 0.4 to 2.5
 
         image_urls = await upload_images_to_comfyapi(cls, images, max_images=4, mime_type="image/png")
         prompt = (

comfy_api_nodes/nodes_ideogram.py

@@ -1,6 +1,6 @@
 from io import BytesIO
 from typing_extensions import override
-from comfy_api.latest import ComfyExtension, IO
+from comfy_api.latest import IO, ComfyExtension
 from PIL import Image
 import numpy as np
 import torch
@@ -11,19 +11,13 @@ from comfy_api_nodes.apis import (
     IdeogramV3Request,
     IdeogramV3EditRequest,
 )
-
-from comfy_api_nodes.apis.client import (
+from comfy_api_nodes.util import (
     ApiEndpoint,
-    HttpMethod,
-    SynchronousOperation,
-)
-
-from comfy_api_nodes.apinode_utils import (
-    download_url_to_bytesio,
     bytesio_to_image_tensor,
+    download_url_as_bytesio,
     resize_mask_to_image,
+    sync_op,
 )
-from server import PromptServer
 
 V1_V1_RES_MAP = {
   "Auto":"AUTO",
@@ -220,7 +214,7 @@ async def download_and_process_images(image_urls):
 
     for image_url in image_urls:
         # Using functions from apinode_utils.py to handle downloading and processing
-        image_bytesio = await download_url_to_bytesio(image_url)  # Download image content to BytesIO
+        image_bytesio = await download_url_as_bytesio(image_url)  # Download image content to BytesIO
         img_tensor = bytesio_to_image_tensor(image_bytesio, mode="RGB")  # Convert to torch.Tensor with RGB mode
         image_tensors.append(img_tensor)
 
@@ -233,19 +227,6 @@ async def download_and_process_images(image_urls):
     return stacked_tensors
 
 
-def display_image_urls_on_node(image_urls, node_id):
-    if node_id and image_urls:
-        if len(image_urls) == 1:
-            PromptServer.instance.send_progress_text(
-                f"Generated Image URL:\n{image_urls[0]}", node_id
-            )
-        else:
-            urls_text = "Generated Image URLs:\n" + "\n".join(
-                f"{i+1}. {url}" for i, url in enumerate(image_urls)
-            )
-            PromptServer.instance.send_progress_text(urls_text, node_id)
-
-
 class IdeogramV1(IO.ComfyNode):
 
     @classmethod
@@ -334,44 +315,30 @@ class IdeogramV1(IO.ComfyNode):
         aspect_ratio = V1_V2_RATIO_MAP.get(aspect_ratio, None)
         model = "V_1_TURBO" if turbo else "V_1"
 
-        auth = {
-            "auth_token": cls.hidden.auth_token_comfy_org,
-            "comfy_api_key": cls.hidden.api_key_comfy_org,
-        }
-        operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path="/proxy/ideogram/generate",
-                method=HttpMethod.POST,
-                request_model=IdeogramGenerateRequest,
-                response_model=IdeogramGenerateResponse,
-            ),
-            request=IdeogramGenerateRequest(
+        response = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/ideogram/generate", method="POST"),
+            response_model=IdeogramGenerateResponse,
+            data=IdeogramGenerateRequest(
                 image_request=ImageRequest(
                     prompt=prompt,
                     model=model,
                     num_images=num_images,
                     seed=seed,
                     aspect_ratio=aspect_ratio if aspect_ratio != "ASPECT_1_1" else None,
-                    magic_prompt_option=(
-                        magic_prompt_option if magic_prompt_option != "AUTO" else None
-                    ),
+                    magic_prompt_option=(magic_prompt_option if magic_prompt_option != "AUTO" else None),
                     negative_prompt=negative_prompt if negative_prompt else None,
                 )
             ),
-            auth_kwargs=auth,
+            max_retries=1,
         )
 
-        response = await operation.execute()
-
         if not response.data or len(response.data) == 0:
             raise Exception("No images were generated in the response")
 
         image_urls = [image_data.url for image_data in response.data if image_data.url]
-
         if not image_urls:
             raise Exception("No image URLs were generated in the response")
-
-        display_image_urls_on_node(image_urls, cls.hidden.unique_id)
         return IO.NodeOutput(await download_and_process_images(image_urls))
 
 
@@ -500,18 +467,11 @@ class IdeogramV2(IO.ComfyNode):
         else:
             final_aspect_ratio = aspect_ratio if aspect_ratio != "ASPECT_1_1" else None
 
-        auth = {
-            "auth_token": cls.hidden.auth_token_comfy_org,
-            "comfy_api_key": cls.hidden.api_key_comfy_org,
-        }
-        operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path="/proxy/ideogram/generate",
-                method=HttpMethod.POST,
-                request_model=IdeogramGenerateRequest,
-                response_model=IdeogramGenerateResponse,
-            ),
-            request=IdeogramGenerateRequest(
+        response = await sync_op(
+            cls,
+            endpoint=ApiEndpoint(path="/proxy/ideogram/generate", method="POST"),
+            response_model=IdeogramGenerateResponse,
+            data=IdeogramGenerateRequest(
                 image_request=ImageRequest(
                     prompt=prompt,
                     model=model,
@@ -519,28 +479,20 @@ class IdeogramV2(IO.ComfyNode):
                     seed=seed,
                     aspect_ratio=final_aspect_ratio,
                     resolution=final_resolution,
-                    magic_prompt_option=(
-                        magic_prompt_option if magic_prompt_option != "AUTO" else None
-                    ),
+                    magic_prompt_option=(magic_prompt_option if magic_prompt_option != "AUTO" else None),
                     style_type=style_type if style_type != "NONE" else None,
                     negative_prompt=negative_prompt if negative_prompt else None,
                     color_palette=color_palette if color_palette else None,
                 )
             ),
-            auth_kwargs=auth,
+            max_retries=1,
         )
-
-        response = await operation.execute()
-
         if not response.data or len(response.data) == 0:
             raise Exception("No images were generated in the response")
 
         image_urls = [image_data.url for image_data in response.data if image_data.url]
-
         if not image_urls:
             raise Exception("No image URLs were generated in the response")
-
-        display_image_urls_on_node(image_urls, cls.hidden.unique_id)
         return IO.NodeOutput(await download_and_process_images(image_urls))
 
 
@@ -656,10 +608,6 @@ class IdeogramV3(IO.ComfyNode):
         character_image=None,
         character_mask=None,
     ):
-        auth = {
-            "auth_token": cls.hidden.auth_token_comfy_org,
-            "comfy_api_key": cls.hidden.api_key_comfy_org,
-        }
         if rendering_speed == "BALANCED":  # for backward compatibility
             rendering_speed = "DEFAULT"
 
@@ -694,9 +642,6 @@ class IdeogramV3(IO.ComfyNode):
 
         # Check if both image and mask are provided for editing mode
         if image is not None and mask is not None:
-            # Edit mode
-            path = "/proxy/ideogram/ideogram-v3/edit"
-
             # Process image and mask
             input_tensor = image.squeeze().cpu()
             # Resize mask to match image dimension
@@ -749,27 +694,20 @@ class IdeogramV3(IO.ComfyNode):
             if character_mask_binary:
                 files["character_mask_binary"] = character_mask_binary
 
-            # Execute the operation for edit mode
-            operation = SynchronousOperation(
-                endpoint=ApiEndpoint(
-                    path=path,
-                    method=HttpMethod.POST,
-                    request_model=IdeogramV3EditRequest,
-                    response_model=IdeogramGenerateResponse,
-                ),
-                request=edit_request,
+            response = await sync_op(
+                cls,
+                ApiEndpoint(path="/proxy/ideogram/ideogram-v3/edit", method="POST"),
+                response_model=IdeogramGenerateResponse,
+                data=edit_request,
                 files=files,
                 content_type="multipart/form-data",
-                auth_kwargs=auth,
+                max_retries=1,
             )
 
         elif image is not None or mask is not None:
             # If only one of image or mask is provided, raise an error
             raise Exception("Ideogram V3 image editing requires both an image AND a mask")
         else:
-            # Generation mode
-            path = "/proxy/ideogram/ideogram-v3/generate"
-
             # Create generation request
             gen_request = IdeogramV3Request(
                 prompt=prompt,
@@ -800,32 +738,22 @@ class IdeogramV3(IO.ComfyNode):
             if files:
                 gen_request.style_type = "AUTO"
 
-            # Execute the operation for generation mode
-            operation = SynchronousOperation(
-                endpoint=ApiEndpoint(
-                    path=path,
-                    method=HttpMethod.POST,
-                    request_model=IdeogramV3Request,
-                    response_model=IdeogramGenerateResponse,
-                ),
-                request=gen_request,
+            response = await sync_op(
+                cls,
+                endpoint=ApiEndpoint(path="/proxy/ideogram/ideogram-v3/generate", method="POST"),
+                response_model=IdeogramGenerateResponse,
+                data=gen_request,
                 files=files if files else None,
                 content_type="multipart/form-data",
-                auth_kwargs=auth,
+                max_retries=1,
             )
 
-        # Execute the operation and process response
-        response = await operation.execute()
-
         if not response.data or len(response.data) == 0:
             raise Exception("No images were generated in the response")
 
         image_urls = [image_data.url for image_data in response.data if image_data.url]
-
         if not image_urls:
             raise Exception("No image URLs were generated in the response")
-
-        display_image_urls_on_node(image_urls, cls.hidden.unique_id)
         return IO.NodeOutput(await download_and_process_images(image_urls))
 
 
@@ -838,5 +766,6 @@ class IdeogramExtension(ComfyExtension):
             IdeogramV3,
         ]
 
+
 async def comfy_entrypoint() -> IdeogramExtension:
     return IdeogramExtension()

comfy_api_nodes/nodes_kling.py

@@ -282,7 +282,7 @@ def validate_input_image(image: torch.Tensor) -> None:
     See: https://app.klingai.com/global/dev/document-api/apiReference/model/imageToVideo
     """
     validate_image_dimensions(image, min_width=300, min_height=300)
-    validate_image_aspect_ratio(image, min_aspect_ratio=1 / 2.5, max_aspect_ratio=2.5)
+    validate_image_aspect_ratio(image, (1, 2.5), (2.5, 1))
 
 
 def get_video_from_response(response) -> KlingVideoResult:

comfy_api_nodes/nodes_ltxv.py

@@ -46,7 +46,7 @@ class TextToVideoNode(IO.ComfyNode):
                     multiline=True,
                     default="",
                 ),
-                IO.Combo.Input("duration", options=[6, 8, 10], default=8),
+                IO.Combo.Input("duration", options=[6, 8, 10, 12, 14, 16, 18, 20], default=8),
                 IO.Combo.Input(
                     "resolution",
                     options=[
@@ -85,6 +85,10 @@ class TextToVideoNode(IO.ComfyNode):
         generate_audio: bool = False,
     ) -> IO.NodeOutput:
         validate_string(prompt, min_length=1, max_length=10000)
+        if duration > 10 and (model != "LTX-2 (Fast)" or resolution != "1920x1080" or fps != 25):
+            raise ValueError(
+                "Durations over 10s are only available for the Fast model at 1920x1080 resolution and 25 FPS."
+            )
         response = await sync_op_raw(
             cls,
             ApiEndpoint("/proxy/ltx/v1/text-to-video", "POST"),
@@ -118,7 +122,7 @@ class ImageToVideoNode(IO.ComfyNode):
                     multiline=True,
                     default="",
                 ),
-                IO.Combo.Input("duration", options=[6, 8, 10], default=8),
+                IO.Combo.Input("duration", options=[6, 8, 10, 12, 14, 16, 18, 20], default=8),
                 IO.Combo.Input(
                     "resolution",
                     options=[
@@ -158,6 +162,10 @@ class ImageToVideoNode(IO.ComfyNode):
         generate_audio: bool = False,
     ) -> IO.NodeOutput:
         validate_string(prompt, min_length=1, max_length=10000)
+        if duration > 10 and (model != "LTX-2 (Fast)" or resolution != "1920x1080" or fps != 25):
+            raise ValueError(
+                "Durations over 10s are only available for the Fast model at 1920x1080 resolution and 25 FPS."
+            )
         if get_number_of_images(image) != 1:
             raise ValueError("Currently only one input image is supported.")
         response = await sync_op_raw(

comfy_api_nodes/nodes_luma.py

@@ -1,69 +1,51 @@
-from __future__ import annotations
-from inspect import cleandoc
 from typing import Optional
+
+import torch
 from typing_extensions import override
-from comfy_api.latest import ComfyExtension, IO
-from comfy_api.input_impl.video_types import VideoFromFile
+
+from comfy_api.latest import IO, ComfyExtension
 from comfy_api_nodes.apis.luma_api import (
-    LumaImageModel,
-    LumaVideoModel,
-    LumaVideoOutputResolution,
-    LumaVideoModelOutputDuration,
     LumaAspectRatio,
-    LumaState,
-    LumaImageGenerationRequest,
-    LumaGenerationRequest,
-    LumaGeneration,
     LumaCharacterRef,
-    LumaModifyImageRef,
+    LumaConceptChain,
+    LumaGeneration,
+    LumaGenerationRequest,
+    LumaImageGenerationRequest,
     LumaImageIdentity,
+    LumaImageModel,
+    LumaImageReference,
+    LumaIO,
+    LumaKeyframes,
+    LumaModifyImageRef,
     LumaReference,
     LumaReferenceChain,
-    LumaImageReference,
-    LumaKeyframes,
-    LumaConceptChain,
-    LumaIO,
+    LumaVideoModel,
+    LumaVideoModelOutputDuration,
+    LumaVideoOutputResolution,
     get_luma_concepts,
 )
-from comfy_api_nodes.apis.client import (
+from comfy_api_nodes.util import (
     ApiEndpoint,
-    HttpMethod,
-    SynchronousOperation,
-    PollingOperation,
-    EmptyRequest,
-)
-from comfy_api_nodes.apinode_utils import (
+    download_url_to_image_tensor,
+    download_url_to_video_output,
+    poll_op,
+    sync_op,
     upload_images_to_comfyapi,
-    process_image_response,
+    validate_string,
 )
-from server import PromptServer
-from comfy_api_nodes.util import validate_string
-
-import aiohttp
-import torch
-from io import BytesIO
 
 LUMA_T2V_AVERAGE_DURATION = 105
 LUMA_I2V_AVERAGE_DURATION = 100
 
-def image_result_url_extractor(response: LumaGeneration):
-    return response.assets.image if hasattr(response, "assets") and hasattr(response.assets, "image") else None
-
-def video_result_url_extractor(response: LumaGeneration):
-    return response.assets.video if hasattr(response, "assets") and hasattr(response.assets, "video") else None
 
 class LumaReferenceNode(IO.ComfyNode):
-    """
-    Holds an image and weight for use with Luma Generate Image node.
-    """
-
     @classmethod
     def define_schema(cls) -> IO.Schema:
         return IO.Schema(
             node_id="LumaReferenceNode",
             display_name="Luma Reference",
             category="api node/image/Luma",
-            description=cleandoc(cls.__doc__ or ""),
+            description="Holds an image and weight for use with Luma Generate Image node.",
             inputs=[
                 IO.Image.Input(
                     "image",
@@ -83,17 +65,10 @@ class LumaReferenceNode(IO.ComfyNode):
                 ),
             ],
             outputs=[IO.Custom(LumaIO.LUMA_REF).Output(display_name="luma_ref")],
-            hidden=[
-                IO.Hidden.auth_token_comfy_org,
-                IO.Hidden.api_key_comfy_org,
-                IO.Hidden.unique_id,
-            ],
         )
 
     @classmethod
-    def execute(
-        cls, image: torch.Tensor, weight: float, luma_ref: LumaReferenceChain = None
-    ) -> IO.NodeOutput:
+    def execute(cls, image: torch.Tensor, weight: float, luma_ref: LumaReferenceChain = None) -> IO.NodeOutput:
         if luma_ref is not None:
             luma_ref = luma_ref.clone()
         else:
@@ -103,17 +78,13 @@ class LumaReferenceNode(IO.ComfyNode):
 
 
 class LumaConceptsNode(IO.ComfyNode):
-    """
-    Holds one or more Camera Concepts for use with Luma Text to Video and Luma Image to Video nodes.
-    """
-
     @classmethod
     def define_schema(cls) -> IO.Schema:
         return IO.Schema(
             node_id="LumaConceptsNode",
             display_name="Luma Concepts",
             category="api node/video/Luma",
-            description=cleandoc(cls.__doc__ or ""),
+            description="Camera Concepts for use with Luma Text to Video and Luma Image to Video nodes.",
             inputs=[
                 IO.Combo.Input(
                     "concept1",
@@ -138,11 +109,6 @@ class LumaConceptsNode(IO.ComfyNode):
                 ),
             ],
             outputs=[IO.Custom(LumaIO.LUMA_CONCEPTS).Output(display_name="luma_concepts")],
-            hidden=[
-                IO.Hidden.auth_token_comfy_org,
-                IO.Hidden.api_key_comfy_org,
-                IO.Hidden.unique_id,
-            ],
         )
 
     @classmethod
@@ -161,17 +127,13 @@ class LumaConceptsNode(IO.ComfyNode):
 
 
 class LumaImageGenerationNode(IO.ComfyNode):
-    """
-    Generates images synchronously based on prompt and aspect ratio.
-    """
-
     @classmethod
     def define_schema(cls) -> IO.Schema:
         return IO.Schema(
             node_id="LumaImageNode",
             display_name="Luma Text to Image",
             category="api node/image/Luma",
-            description=cleandoc(cls.__doc__ or ""),
+            description="Generates images synchronously based on prompt and aspect ratio.",
             inputs=[
                 IO.String.Input(
                     "prompt",
@@ -237,45 +199,30 @@ class LumaImageGenerationNode(IO.ComfyNode):
         aspect_ratio: str,
         seed,
         style_image_weight: float,
-        image_luma_ref: LumaReferenceChain = None,
-        style_image: torch.Tensor = None,
-        character_image: torch.Tensor = None,
+        image_luma_ref: Optional[LumaReferenceChain] = None,
+        style_image: Optional[torch.Tensor] = None,
+        character_image: Optional[torch.Tensor] = None,
     ) -> IO.NodeOutput:
         validate_string(prompt, strip_whitespace=True, min_length=3)
-        auth_kwargs = {
-            "auth_token": cls.hidden.auth_token_comfy_org,
-            "comfy_api_key": cls.hidden.api_key_comfy_org,
-        }
         # handle image_luma_ref
         api_image_ref = None
         if image_luma_ref is not None:
-            api_image_ref = await cls._convert_luma_refs(
-                image_luma_ref, max_refs=4, auth_kwargs=auth_kwargs,
-            )
+            api_image_ref = await cls._convert_luma_refs(image_luma_ref, max_refs=4)
         # handle style_luma_ref
         api_style_ref = None
         if style_image is not None:
-            api_style_ref = await cls._convert_style_image(
-                style_image, weight=style_image_weight, auth_kwargs=auth_kwargs,
-            )
+            api_style_ref = await cls._convert_style_image(style_image, weight=style_image_weight)
         # handle character_ref images
         character_ref = None
         if character_image is not None:
-            download_urls = await upload_images_to_comfyapi(
-                character_image, max_images=4, auth_kwargs=auth_kwargs,
-            )
-            character_ref = LumaCharacterRef(
-                identity0=LumaImageIdentity(images=download_urls)
-            )
+            download_urls = await upload_images_to_comfyapi(cls, character_image, max_images=4)
+            character_ref = LumaCharacterRef(identity0=LumaImageIdentity(images=download_urls))
 
-        operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path="/proxy/luma/generations/image",
-                method=HttpMethod.POST,
-                request_model=LumaImageGenerationRequest,
-                response_model=LumaGeneration,
-            ),
-            request=LumaImageGenerationRequest(
+        response_api = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/luma/generations/image", method="POST"),
+            response_model=LumaGeneration,
+            data=LumaImageGenerationRequest(
                 prompt=prompt,
                 model=model,
                 aspect_ratio=aspect_ratio,
@@ -283,41 +230,21 @@ class LumaImageGenerationNode(IO.ComfyNode):
                 style_ref=api_style_ref,
                 character_ref=character_ref,
             ),
-            auth_kwargs=auth_kwargs,
         )
-        response_api: LumaGeneration = await operation.execute()
-
-        operation = PollingOperation(
-            poll_endpoint=ApiEndpoint(
-                path=f"/proxy/luma/generations/{response_api.id}",
-                method=HttpMethod.GET,
-                request_model=EmptyRequest,
-                response_model=LumaGeneration,
-            ),
-            completed_statuses=[LumaState.completed],
-            failed_statuses=[LumaState.failed],
+        response_poll = await poll_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/luma/generations/{response_api.id}"),
+            response_model=LumaGeneration,
             status_extractor=lambda x: x.state,
-            result_url_extractor=image_result_url_extractor,
-            node_id=cls.hidden.unique_id,
-            auth_kwargs=auth_kwargs,
         )
-        response_poll = await operation.execute()
-
-        async with aiohttp.ClientSession() as session:
-            async with session.get(response_poll.assets.image) as img_response:
-                img = process_image_response(await img_response.content.read())
-        return IO.NodeOutput(img)
+        return IO.NodeOutput(await download_url_to_image_tensor(response_poll.assets.image))
 
     @classmethod
-    async def _convert_luma_refs(
-        cls, luma_ref: LumaReferenceChain, max_refs: int, auth_kwargs: Optional[dict[str,str]] = None
-    ):
+    async def _convert_luma_refs(cls, luma_ref: LumaReferenceChain, max_refs: int):
         luma_urls = []
         ref_count = 0
         for ref in luma_ref.refs:
-            download_urls = await upload_images_to_comfyapi(
-                ref.image, max_images=1, auth_kwargs=auth_kwargs
-            )
+            download_urls = await upload_images_to_comfyapi(cls, ref.image, max_images=1)
             luma_urls.append(download_urls[0])
             ref_count += 1
             if ref_count >= max_refs:
@@ -325,27 +252,19 @@ class LumaImageGenerationNode(IO.ComfyNode):
         return luma_ref.create_api_model(download_urls=luma_urls, max_refs=max_refs)
 
     @classmethod
-    async def _convert_style_image(
-        cls, style_image: torch.Tensor, weight: float, auth_kwargs: Optional[dict[str,str]] = None
-    ):
-        chain = LumaReferenceChain(
-            first_ref=LumaReference(image=style_image, weight=weight)
-        )
-        return await cls._convert_luma_refs(chain, max_refs=1, auth_kwargs=auth_kwargs)
+    async def _convert_style_image(cls, style_image: torch.Tensor, weight: float):
+        chain = LumaReferenceChain(first_ref=LumaReference(image=style_image, weight=weight))
+        return await cls._convert_luma_refs(chain, max_refs=1)
 
 
 class LumaImageModifyNode(IO.ComfyNode):
-    """
-    Modifies images synchronously based on prompt and aspect ratio.
-    """
-
     @classmethod
     def define_schema(cls) -> IO.Schema:
         return IO.Schema(
             node_id="LumaImageModifyNode",
             display_name="Luma Image to Image",
             category="api node/image/Luma",
-            description=cleandoc(cls.__doc__ or ""),
+            description="Modifies images synchronously based on prompt and aspect ratio.",
             inputs=[
                 IO.Image.Input(
                     "image",
@@ -395,68 +314,37 @@ class LumaImageModifyNode(IO.ComfyNode):
         image_weight: float,
         seed,
     ) -> IO.NodeOutput:
-        auth_kwargs = {
-            "auth_token": cls.hidden.auth_token_comfy_org,
-            "comfy_api_key": cls.hidden.api_key_comfy_org,
-        }
-        # first, upload image
-        download_urls = await upload_images_to_comfyapi(
-            image, max_images=1, auth_kwargs=auth_kwargs,
-        )
+        download_urls = await upload_images_to_comfyapi(cls, image, max_images=1)
         image_url = download_urls[0]
-        # next, make Luma call with download url provided
-        operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path="/proxy/luma/generations/image",
-                method=HttpMethod.POST,
-                request_model=LumaImageGenerationRequest,
-                response_model=LumaGeneration,
-            ),
-            request=LumaImageGenerationRequest(
+        response_api = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/luma/generations/image", method="POST"),
+            response_model=LumaGeneration,
+            data=LumaImageGenerationRequest(
                 prompt=prompt,
                 model=model,
                 modify_image_ref=LumaModifyImageRef(
-                    url=image_url, weight=round(max(min(1.0-image_weight, 0.98), 0.0), 2)
+                    url=image_url, weight=round(max(min(1.0 - image_weight, 0.98), 0.0), 2)
                 ),
             ),
-            auth_kwargs=auth_kwargs,
         )
-        response_api: LumaGeneration = await operation.execute()
-
-        operation = PollingOperation(
-            poll_endpoint=ApiEndpoint(
-                path=f"/proxy/luma/generations/{response_api.id}",
-                method=HttpMethod.GET,
-                request_model=EmptyRequest,
-                response_model=LumaGeneration,
-            ),
-            completed_statuses=[LumaState.completed],
-            failed_statuses=[LumaState.failed],
+        response_poll = await poll_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/luma/generations/{response_api.id}"),
+            response_model=LumaGeneration,
             status_extractor=lambda x: x.state,
-            result_url_extractor=image_result_url_extractor,
-            node_id=cls.hidden.unique_id,
-            auth_kwargs=auth_kwargs,
         )
-        response_poll = await operation.execute()
-
-        async with aiohttp.ClientSession() as session:
-            async with session.get(response_poll.assets.image) as img_response:
-                img = process_image_response(await img_response.content.read())
-        return IO.NodeOutput(img)
+        return IO.NodeOutput(await download_url_to_image_tensor(response_poll.assets.image))
 
 
 class LumaTextToVideoGenerationNode(IO.ComfyNode):
-    """
-    Generates videos synchronously based on prompt and output_size.
-    """
-
     @classmethod
     def define_schema(cls) -> IO.Schema:
         return IO.Schema(
             node_id="LumaVideoNode",
             display_name="Luma Text to Video",
             category="api node/video/Luma",
-            description=cleandoc(cls.__doc__ or ""),
+            description="Generates videos synchronously based on prompt and output_size.",
             inputs=[
                 IO.String.Input(
                     "prompt",
@@ -498,7 +386,7 @@ class LumaTextToVideoGenerationNode(IO.ComfyNode):
                     "luma_concepts",
                     tooltip="Optional Camera Concepts to dictate camera motion via the Luma Concepts node.",
                     optional=True,
-                )
+                ),
             ],
             outputs=[IO.Video.Output()],
             hidden=[
@@ -519,24 +407,17 @@ class LumaTextToVideoGenerationNode(IO.ComfyNode):
         duration: str,
         loop: bool,
         seed,
-        luma_concepts: LumaConceptChain = None,
+        luma_concepts: Optional[LumaConceptChain] = None,
     ) -> IO.NodeOutput:
         validate_string(prompt, strip_whitespace=False, min_length=3)
         duration = duration if model != LumaVideoModel.ray_1_6 else None
         resolution = resolution if model != LumaVideoModel.ray_1_6 else None
 
-        auth_kwargs = {
-            "auth_token": cls.hidden.auth_token_comfy_org,
-            "comfy_api_key": cls.hidden.api_key_comfy_org,
-        }
-        operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path="/proxy/luma/generations",
-                method=HttpMethod.POST,
-                request_model=LumaGenerationRequest,
-                response_model=LumaGeneration,
-            ),
-            request=LumaGenerationRequest(
+        response_api = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/luma/generations", method="POST"),
+            response_model=LumaGeneration,
+            data=LumaGenerationRequest(
                 prompt=prompt,
                 model=model,
                 resolution=resolution,
@@ -545,47 +426,25 @@ class LumaTextToVideoGenerationNode(IO.ComfyNode):
                 loop=loop,
                 concepts=luma_concepts.create_api_model() if luma_concepts else None,
             ),
-            auth_kwargs=auth_kwargs,
         )
-        response_api: LumaGeneration = await operation.execute()
-
-        if cls.hidden.unique_id:
-            PromptServer.instance.send_progress_text(f"Luma video generation started: {response_api.id}", cls.hidden.unique_id)
-
-        operation = PollingOperation(
-            poll_endpoint=ApiEndpoint(
-                path=f"/proxy/luma/generations/{response_api.id}",
-                method=HttpMethod.GET,
-                request_model=EmptyRequest,
-                response_model=LumaGeneration,
-            ),
-            completed_statuses=[LumaState.completed],
-            failed_statuses=[LumaState.failed],
+        response_poll = await poll_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/luma/generations/{response_api.id}"),
+            response_model=LumaGeneration,
             status_extractor=lambda x: x.state,
-            result_url_extractor=video_result_url_extractor,
-            node_id=cls.hidden.unique_id,
             estimated_duration=LUMA_T2V_AVERAGE_DURATION,
-            auth_kwargs=auth_kwargs,
         )
-        response_poll = await operation.execute()
-
-        async with aiohttp.ClientSession() as session:
-            async with session.get(response_poll.assets.video) as vid_response:
-                return IO.NodeOutput(VideoFromFile(BytesIO(await vid_response.content.read())))
+        return IO.NodeOutput(await download_url_to_video_output(response_poll.assets.video))
 
 
 class LumaImageToVideoGenerationNode(IO.ComfyNode):
-    """
-    Generates videos synchronously based on prompt, input images, and output_size.
-    """
-
     @classmethod
     def define_schema(cls) -> IO.Schema:
         return IO.Schema(
             node_id="LumaImageToVideoNode",
             display_name="Luma Image to Video",
             category="api node/video/Luma",
-            description=cleandoc(cls.__doc__ or ""),
+            description="Generates videos synchronously based on prompt, input images, and output_size.",
             inputs=[
                 IO.String.Input(
                     "prompt",
@@ -637,7 +496,7 @@ class LumaImageToVideoGenerationNode(IO.ComfyNode):
                     "luma_concepts",
                     tooltip="Optional Camera Concepts to dictate camera motion via the Luma Concepts node.",
                     optional=True,
-                )
+                ),
             ],
             outputs=[IO.Video.Output()],
             hidden=[
@@ -662,25 +521,15 @@ class LumaImageToVideoGenerationNode(IO.ComfyNode):
         luma_concepts: LumaConceptChain = None,
     ) -> IO.NodeOutput:
         if first_image is None and last_image is None:
-            raise Exception(
-                "At least one of first_image and last_image requires an input."
-            )
-        auth_kwargs = {
-            "auth_token": cls.hidden.auth_token_comfy_org,
-            "comfy_api_key": cls.hidden.api_key_comfy_org,
-        }
-        keyframes = await cls._convert_to_keyframes(first_image, last_image, auth_kwargs=auth_kwargs)
+            raise Exception("At least one of first_image and last_image requires an input.")
+        keyframes = await cls._convert_to_keyframes(first_image, last_image)
         duration = duration if model != LumaVideoModel.ray_1_6 else None
         resolution = resolution if model != LumaVideoModel.ray_1_6 else None
-
-        operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path="/proxy/luma/generations",
-                method=HttpMethod.POST,
-                request_model=LumaGenerationRequest,
-                response_model=LumaGeneration,
-            ),
-            request=LumaGenerationRequest(
+        response_api = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/luma/generations", method="POST"),
+            response_model=LumaGeneration,
+            data=LumaGenerationRequest(
                 prompt=prompt,
                 model=model,
                 aspect_ratio=LumaAspectRatio.ratio_16_9,  # ignored, but still needed by the API for some reason
@@ -690,54 +539,31 @@ class LumaImageToVideoGenerationNode(IO.ComfyNode):
                 keyframes=keyframes,
                 concepts=luma_concepts.create_api_model() if luma_concepts else None,
             ),
-            auth_kwargs=auth_kwargs,
         )
-        response_api: LumaGeneration = await operation.execute()
-
-        if cls.hidden.unique_id:
-            PromptServer.instance.send_progress_text(f"Luma video generation started: {response_api.id}", cls.hidden.unique_id)
-
-        operation = PollingOperation(
-            poll_endpoint=ApiEndpoint(
-                path=f"/proxy/luma/generations/{response_api.id}",
-                method=HttpMethod.GET,
-                request_model=EmptyRequest,
-                response_model=LumaGeneration,
-            ),
-            completed_statuses=[LumaState.completed],
-            failed_statuses=[LumaState.failed],
+        response_poll = await poll_op(
+            cls,
+            poll_endpoint=ApiEndpoint(path=f"/proxy/luma/generations/{response_api.id}"),
+            response_model=LumaGeneration,
             status_extractor=lambda x: x.state,
-            result_url_extractor=video_result_url_extractor,
-            node_id=cls.hidden.unique_id,
             estimated_duration=LUMA_I2V_AVERAGE_DURATION,
-            auth_kwargs=auth_kwargs,
         )
-        response_poll = await operation.execute()
-
-        async with aiohttp.ClientSession() as session:
-            async with session.get(response_poll.assets.video) as vid_response:
-                return IO.NodeOutput(VideoFromFile(BytesIO(await vid_response.content.read())))
+        return IO.NodeOutput(await download_url_to_video_output(response_poll.assets.video))
 
     @classmethod
     async def _convert_to_keyframes(
         cls,
         first_image: torch.Tensor = None,
         last_image: torch.Tensor = None,
-        auth_kwargs: Optional[dict[str,str]] = None,
     ):
         if first_image is None and last_image is None:
             return None
         frame0 = None
         frame1 = None
         if first_image is not None:
-            download_urls = await upload_images_to_comfyapi(
-                first_image, max_images=1, auth_kwargs=auth_kwargs,
-            )
+            download_urls = await upload_images_to_comfyapi(cls, first_image, max_images=1)
             frame0 = LumaImageReference(type="image", url=download_urls[0])
         if last_image is not None:
-            download_urls = await upload_images_to_comfyapi(
-                last_image, max_images=1, auth_kwargs=auth_kwargs,
-            )
+            download_urls = await upload_images_to_comfyapi(cls, last_image, max_images=1)
             frame1 = LumaImageReference(type="image", url=download_urls[0])
         return LumaKeyframes(frame0=frame0, frame1=frame1)
 

comfy_api_nodes/nodes_minimax.py

@@ -1,71 +1,57 @@
-from inspect import cleandoc
 from typing import Optional
-import logging
-import torch
 
+import torch
 from typing_extensions import override
-from comfy_api.latest import ComfyExtension, IO
-from comfy_api.input_impl.video_types import VideoFromFile
-from comfy_api_nodes.apis import (
+
+from comfy_api.latest import IO, ComfyExtension
+from comfy_api_nodes.apis.minimax_api import (
+    MinimaxFileRetrieveResponse,
+    MiniMaxModel,
+    MinimaxTaskResultResponse,
     MinimaxVideoGenerationRequest,
     MinimaxVideoGenerationResponse,
-    MinimaxFileRetrieveResponse,
-    MinimaxTaskResultResponse,
     SubjectReferenceItem,
-    MiniMaxModel,
 )
-from comfy_api_nodes.apis.client import (
+from comfy_api_nodes.util import (
     ApiEndpoint,
-    HttpMethod,
-    SynchronousOperation,
-    PollingOperation,
-    EmptyRequest,
-)
-from comfy_api_nodes.apinode_utils import (
-    download_url_to_bytesio,
+    download_url_to_video_output,
+    poll_op,
+    sync_op,
     upload_images_to_comfyapi,
+    validate_string,
 )
-from comfy_api_nodes.util import validate_string
-from server import PromptServer
-
 
 I2V_AVERAGE_DURATION = 114
 T2V_AVERAGE_DURATION = 234
 
 
 async def _generate_mm_video(
+    cls: type[IO.ComfyNode],
     *,
-    auth: dict[str, str],
-    node_id: str,
     prompt_text: str,
     seed: int,
     model: str,
-    image: Optional[torch.Tensor] = None,   # used for ImageToVideo
-    subject: Optional[torch.Tensor] = None, # used for SubjectToVideo
+    image: Optional[torch.Tensor] = None,  # used for ImageToVideo
+    subject: Optional[torch.Tensor] = None,  # used for SubjectToVideo
     average_duration: Optional[int] = None,
 ) -> IO.NodeOutput:
     if image is None:
         validate_string(prompt_text, field_name="prompt_text")
-    # upload image, if passed in
     image_url = None
     if image is not None:
-        image_url = (await upload_images_to_comfyapi(image, max_images=1, auth_kwargs=auth))[0]
+        image_url = (await upload_images_to_comfyapi(cls, image, max_images=1))[0]
 
     # TODO: figure out how to deal with subject properly, API returns invalid params when using S2V-01 model
     subject_reference = None
     if subject is not None:
-        subject_url = (await upload_images_to_comfyapi(subject, max_images=1, auth_kwargs=auth))[0]
+        subject_url = (await upload_images_to_comfyapi(cls, subject, max_images=1))[0]
         subject_reference = [SubjectReferenceItem(image=subject_url)]
 
-
-    video_generate_operation = SynchronousOperation(
-        endpoint=ApiEndpoint(
-            path="/proxy/minimax/video_generation",
-            method=HttpMethod.POST,
-            request_model=MinimaxVideoGenerationRequest,
-            response_model=MinimaxVideoGenerationResponse,
-        ),
-        request=MinimaxVideoGenerationRequest(
+    response = await sync_op(
+        cls,
+        ApiEndpoint(path="/proxy/minimax/video_generation", method="POST"),
+        response_model=MinimaxVideoGenerationResponse,
+        data=MinimaxVideoGenerationRequest(
             model=MiniMaxModel(model),
             prompt=prompt_text,
             callback_url=None,
@@ -73,81 +59,50 @@ async def _generate_mm_video(
             subject_reference=subject_reference,
             prompt_optimizer=None,
         ),
-        auth_kwargs=auth,
     )
-    response = await video_generate_operation.execute()
 
     task_id = response.task_id
     if not task_id:
         raise Exception(f"MiniMax generation failed: {response.base_resp}")
 
-    video_generate_operation = PollingOperation(
-        poll_endpoint=ApiEndpoint(
-            path="/proxy/minimax/query/video_generation",
-            method=HttpMethod.GET,
-            request_model=EmptyRequest,
-            response_model=MinimaxTaskResultResponse,
-            query_params={"task_id": task_id},
-        ),
-        completed_statuses=["Success"],
-        failed_statuses=["Fail"],
+    task_result = await poll_op(
+        cls,
+        ApiEndpoint(path="/proxy/minimax/query/video_generation", query_params={"task_id": task_id}),
+        response_model=MinimaxTaskResultResponse,
         status_extractor=lambda x: x.status.value,
         estimated_duration=average_duration,
-        node_id=node_id,
-        auth_kwargs=auth,
     )
-    task_result = await video_generate_operation.execute()
 
     file_id = task_result.file_id
     if file_id is None:
         raise Exception("Request was not successful. Missing file ID.")
-    file_retrieve_operation = SynchronousOperation(
-        endpoint=ApiEndpoint(
-            path="/proxy/minimax/files/retrieve",
-            method=HttpMethod.GET,
-            request_model=EmptyRequest,
-            response_model=MinimaxFileRetrieveResponse,
-            query_params={"file_id": int(file_id)},
-        ),
-        request=EmptyRequest(),
-        auth_kwargs=auth,
+    file_result = await sync_op(
+        cls,
+        ApiEndpoint(path="/proxy/minimax/files/retrieve", query_params={"file_id": int(file_id)}),
+        response_model=MinimaxFileRetrieveResponse,
     )
-    file_result = await file_retrieve_operation.execute()
 
     file_url = file_result.file.download_url
     if file_url is None:
-        raise Exception(
-            f"No video was found in the response. Full response: {file_result.model_dump()}"
-        )
-    logging.info("Generated video URL: %s", file_url)
-    if node_id:
-        if hasattr(file_result.file, "backup_download_url"):
-            message = f"Result URL: {file_url}\nBackup URL: {file_result.file.backup_download_url}"
-        else:
-            message = f"Result URL: {file_url}"
-        PromptServer.instance.send_progress_text(message, node_id)
-
-    # Download and return as VideoFromFile
-    video_io = await download_url_to_bytesio(file_url)
-    if video_io is None:
-        error_msg = f"Failed to download video from {file_url}"
-        logging.error(error_msg)
-        raise Exception(error_msg)
-    return IO.NodeOutput(VideoFromFile(video_io))
+        raise Exception(f"No video was found in the response. Full response: {file_result.model_dump()}")
+    if file_result.file.backup_download_url:
+        try:
+            return IO.NodeOutput(await download_url_to_video_output(file_url, timeout=10, max_retries=2))
+        except Exception:  # if we have a second URL to retrieve the result, try again using that one
+            return IO.NodeOutput(
+                await download_url_to_video_output(file_result.file.backup_download_url, max_retries=3)
+            )
+    return IO.NodeOutput(await download_url_to_video_output(file_url))
 
 
 class MinimaxTextToVideoNode(IO.ComfyNode):
-    """
-    Generates videos synchronously based on a prompt, and optional parameters using MiniMax's API.
-    """
-
     @classmethod
     def define_schema(cls) -> IO.Schema:
         return IO.Schema(
             node_id="MinimaxTextToVideoNode",
             display_name="MiniMax Text to Video",
             category="api node/video/MiniMax",
-            description=cleandoc(cls.__doc__ or ""),
+            description="Generates videos synchronously based on a prompt, and optional parameters.",
             inputs=[
                 IO.String.Input(
                     "prompt_text",
@@ -189,11 +144,7 @@ class MinimaxTextToVideoNode(IO.ComfyNode):
         seed: int = 0,
     ) -> IO.NodeOutput:
         return await _generate_mm_video(
-            auth={
-                "auth_token": cls.hidden.auth_token_comfy_org,
-                "comfy_api_key": cls.hidden.api_key_comfy_org,
-            },
-            node_id=cls.hidden.unique_id,
+            cls,
             prompt_text=prompt_text,
             seed=seed,
             model=model,
@@ -204,17 +155,13 @@ class MinimaxTextToVideoNode(IO.ComfyNode):
 
 
 class MinimaxImageToVideoNode(IO.ComfyNode):
-    """
-    Generates videos synchronously based on an image and prompt, and optional parameters using MiniMax's API.
-    """
-
     @classmethod
     def define_schema(cls) -> IO.Schema:
         return IO.Schema(
             node_id="MinimaxImageToVideoNode",
             display_name="MiniMax Image to Video",
             category="api node/video/MiniMax",
-            description=cleandoc(cls.__doc__ or ""),
+            description="Generates videos synchronously based on an image and prompt, and optional parameters.",
             inputs=[
                 IO.Image.Input(
                     "image",
@@ -261,11 +208,7 @@ class MinimaxImageToVideoNode(IO.ComfyNode):
         seed: int = 0,
     ) -> IO.NodeOutput:
         return await _generate_mm_video(
-            auth={
-                "auth_token": cls.hidden.auth_token_comfy_org,
-                "comfy_api_key": cls.hidden.api_key_comfy_org,
-            },
-            node_id=cls.hidden.unique_id,
+            cls,
             prompt_text=prompt_text,
             seed=seed,
             model=model,
@@ -276,17 +219,13 @@ class MinimaxImageToVideoNode(IO.ComfyNode):
 
 
 class MinimaxSubjectToVideoNode(IO.ComfyNode):
-    """
-    Generates videos synchronously based on an image and prompt, and optional parameters using MiniMax's API.
-    """
-
     @classmethod
     def define_schema(cls) -> IO.Schema:
         return IO.Schema(
             node_id="MinimaxSubjectToVideoNode",
             display_name="MiniMax Subject to Video",
             category="api node/video/MiniMax",
-            description=cleandoc(cls.__doc__ or ""),
+            description="Generates videos synchronously based on an image and prompt, and optional parameters.",
             inputs=[
                 IO.Image.Input(
                     "subject",
@@ -333,11 +272,7 @@ class MinimaxSubjectToVideoNode(IO.ComfyNode):
         seed: int = 0,
     ) -> IO.NodeOutput:
         return await _generate_mm_video(
-            auth={
-                "auth_token": cls.hidden.auth_token_comfy_org,
-                "comfy_api_key": cls.hidden.api_key_comfy_org,
-            },
-            node_id=cls.hidden.unique_id,
+            cls,
             prompt_text=prompt_text,
             seed=seed,
             model=model,
@@ -348,15 +283,13 @@ class MinimaxSubjectToVideoNode(IO.ComfyNode):
 
 
 class MinimaxHailuoVideoNode(IO.ComfyNode):
-    """Generates videos from prompt, with optional start frame using the new MiniMax Hailuo-02 model."""
-
     @classmethod
     def define_schema(cls) -> IO.Schema:
         return IO.Schema(
             node_id="MinimaxHailuoVideoNode",
             display_name="MiniMax Hailuo Video",
             category="api node/video/MiniMax",
-            description=cleandoc(cls.__doc__ or ""),
+            description="Generates videos from prompt, with optional start frame using the new MiniMax Hailuo-02 model.",
             inputs=[
                 IO.String.Input(
                     "prompt_text",
@@ -420,10 +353,6 @@ class MinimaxHailuoVideoNode(IO.ComfyNode):
         resolution: str = "768P",
         model: str = "MiniMax-Hailuo-02",
     ) -> IO.NodeOutput:
-        auth = {
-            "auth_token": cls.hidden.auth_token_comfy_org,
-            "comfy_api_key": cls.hidden.api_key_comfy_org,
-        }
         if first_frame_image is None:
             validate_string(prompt_text, field_name="prompt_text")
 
@@ -435,16 +364,13 @@ class MinimaxHailuoVideoNode(IO.ComfyNode):
         # upload image, if passed in
         image_url = None
         if first_frame_image is not None:
-            image_url = (await upload_images_to_comfyapi(first_frame_image, max_images=1, auth_kwargs=auth))[0]
+            image_url = (await upload_images_to_comfyapi(cls, first_frame_image, max_images=1))[0]
 
-        video_generate_operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path="/proxy/minimax/video_generation",
-                method=HttpMethod.POST,
-                request_model=MinimaxVideoGenerationRequest,
-                response_model=MinimaxVideoGenerationResponse,
-            ),
-            request=MinimaxVideoGenerationRequest(
+        response = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/minimax/video_generation", method="POST"),
+            response_model=MinimaxVideoGenerationResponse,
+            data=MinimaxVideoGenerationRequest(
                 model=MiniMaxModel(model),
                 prompt=prompt_text,
                 callback_url=None,
@@ -453,67 +379,42 @@ class MinimaxHailuoVideoNode(IO.ComfyNode):
                 duration=duration,
                 resolution=resolution,
             ),
-            auth_kwargs=auth,
         )
-        response = await video_generate_operation.execute()
 
         task_id = response.task_id
         if not task_id:
             raise Exception(f"MiniMax generation failed: {response.base_resp}")
 
         average_duration = 120 if resolution == "768P" else 240
-        video_generate_operation = PollingOperation(
-            poll_endpoint=ApiEndpoint(
-                path="/proxy/minimax/query/video_generation",
-                method=HttpMethod.GET,
-                request_model=EmptyRequest,
-                response_model=MinimaxTaskResultResponse,
-                query_params={"task_id": task_id},
-            ),
-            completed_statuses=["Success"],
-            failed_statuses=["Fail"],
+        task_result = await poll_op(
+            cls,
+            ApiEndpoint(path="/proxy/minimax/query/video_generation", query_params={"task_id": task_id}),
+            response_model=MinimaxTaskResultResponse,
             status_extractor=lambda x: x.status.value,
             estimated_duration=average_duration,
-            node_id=cls.hidden.unique_id,
-            auth_kwargs=auth,
         )
-        task_result = await video_generate_operation.execute()
 
         file_id = task_result.file_id
         if file_id is None:
             raise Exception("Request was not successful. Missing file ID.")
-        file_retrieve_operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path="/proxy/minimax/files/retrieve",
-                method=HttpMethod.GET,
-                request_model=EmptyRequest,
-                response_model=MinimaxFileRetrieveResponse,
-                query_params={"file_id": int(file_id)},
-            ),
-            request=EmptyRequest(),
-            auth_kwargs=auth,
+        file_result = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/minimax/files/retrieve", query_params={"file_id": int(file_id)}),
+            response_model=MinimaxFileRetrieveResponse,
         )
-        file_result = await file_retrieve_operation.execute()
 
         file_url = file_result.file.download_url
         if file_url is None:
-            raise Exception(
-                f"No video was found in the response. Full response: {file_result.model_dump()}"
-            )
-        logging.info("Generated video URL: %s", file_url)
-        if cls.hidden.unique_id:
-            if hasattr(file_result.file, "backup_download_url"):
-                message = f"Result URL: {file_url}\nBackup URL: {file_result.file.backup_download_url}"
-            else:
-                message = f"Result URL: {file_url}"
-            PromptServer.instance.send_progress_text(message, cls.hidden.unique_id)
+            raise Exception(f"No video was found in the response. Full response: {file_result.model_dump()}")
 
-        video_io = await download_url_to_bytesio(file_url)
-        if video_io is None:
-            error_msg = f"Failed to download video from {file_url}"
-            logging.error(error_msg)
-            raise Exception(error_msg)
-        return IO.NodeOutput(VideoFromFile(video_io))
+        if file_result.file.backup_download_url:
+            try:
+                return IO.NodeOutput(await download_url_to_video_output(file_url, timeout=10, max_retries=2))
+            except Exception:  # if we have a second URL to retrieve the result, try again using that one
+                return IO.NodeOutput(
+                    await download_url_to_video_output(file_result.file.backup_download_url, max_retries=3)
+                )
+        return IO.NodeOutput(await download_url_to_video_output(file_url))
 
 
 class MinimaxExtension(ComfyExtension):

comfy_api_nodes/nodes_pika.py

@@ -7,24 +7,23 @@ from __future__ import annotations
 
 from io import BytesIO
 import logging
-from typing import Optional, TypeVar
+from typing import Optional
 
 import torch
 
 from typing_extensions import override
 from comfy_api.latest import ComfyExtension, IO
 from comfy_api.input_impl.video_types import VideoCodec, VideoContainer, VideoInput
-from comfy_api_nodes.apis import pika_defs
-from comfy_api_nodes.apis.client import (
+from comfy_api_nodes.apis import pika_api as pika_defs
+from comfy_api_nodes.util import (
+    validate_string,
+    download_url_to_video_output,
+    tensor_to_bytesio,
     ApiEndpoint,
-    EmptyRequest,
-    HttpMethod,
-    PollingOperation,
-    SynchronousOperation,
+    sync_op,
+    poll_op,
 )
-from comfy_api_nodes.util import validate_string, download_url_to_video_output, tensor_to_bytesio
 
-R = TypeVar("R")
 
 PATH_PIKADDITIONS = "/proxy/pika/generate/pikadditions"
 PATH_PIKASWAPS = "/proxy/pika/generate/pikaswaps"
@@ -40,28 +39,18 @@ PATH_VIDEO_GET = "/proxy/pika/videos"
 
 
 async def execute_task(
-    initial_operation: SynchronousOperation[R, pika_defs.PikaGenerateResponse],
-    auth_kwargs: Optional[dict[str, str]] = None,
-    node_id: Optional[str] = None,
+    task_id: str,
+    cls: type[IO.ComfyNode],
 ) -> IO.NodeOutput:
-    task_id = (await initial_operation.execute()).video_id
-    final_response: pika_defs.PikaVideoResponse = await PollingOperation(
-        poll_endpoint=ApiEndpoint(
-            path=f"{PATH_VIDEO_GET}/{task_id}",
-            method=HttpMethod.GET,
-            request_model=EmptyRequest,
-            response_model=pika_defs.PikaVideoResponse,
-        ),
-        completed_statuses=["finished"],
-        failed_statuses=["failed", "cancelled"],
+    final_response: pika_defs.PikaVideoResponse = await poll_op(
+        cls,
+        ApiEndpoint(path=f"{PATH_VIDEO_GET}/{task_id}"),
+        response_model=pika_defs.PikaVideoResponse,
         status_extractor=lambda response: (response.status.value if response.status else None),
         progress_extractor=lambda response: (response.progress if hasattr(response, "progress") else None),
-        auth_kwargs=auth_kwargs,
-        result_url_extractor=lambda response: (response.url if hasattr(response, "url") else None),
-        node_id=node_id,
         estimated_duration=60,
         max_poll_attempts=240,
-    ).execute()
+    )
     if not final_response.url:
         error_msg = f"Pika task {task_id} succeeded but no video data found in response:\n{final_response}"
         logging.error(error_msg)
@@ -124,23 +113,15 @@ class PikaImageToVideo(IO.ComfyNode):
             resolution=resolution,
             duration=duration,
         )
-        auth = {
-            "auth_token": cls.hidden.auth_token_comfy_org,
-            "comfy_api_key": cls.hidden.api_key_comfy_org,
-        }
-        initial_operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path=PATH_IMAGE_TO_VIDEO,
-                method=HttpMethod.POST,
-                request_model=pika_defs.PikaBodyGenerate22I2vGenerate22I2vPost,
-                response_model=pika_defs.PikaGenerateResponse,
-            ),
-            request=pika_request_data,
+        initial_operation = await sync_op(
+            cls,
+            ApiEndpoint(path=PATH_IMAGE_TO_VIDEO, method="POST"),
+            response_model=pika_defs.PikaGenerateResponse,
+            data=pika_request_data,
             files=pika_files,
             content_type="multipart/form-data",
-            auth_kwargs=auth,
         )
-        return await execute_task(initial_operation, auth_kwargs=auth, node_id=cls.hidden.unique_id)
+        return await execute_task(initial_operation.video_id, cls)
 
 
 class PikaTextToVideoNode(IO.ComfyNode):
@@ -183,18 +164,11 @@ class PikaTextToVideoNode(IO.ComfyNode):
         duration: int,
         aspect_ratio: float,
     ) -> IO.NodeOutput:
-        auth = {
-            "auth_token": cls.hidden.auth_token_comfy_org,
-            "comfy_api_key": cls.hidden.api_key_comfy_org,
-        }
-        initial_operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path=PATH_TEXT_TO_VIDEO,
-                method=HttpMethod.POST,
-                request_model=pika_defs.PikaBodyGenerate22T2vGenerate22T2vPost,
-                response_model=pika_defs.PikaGenerateResponse,
-            ),
-            request=pika_defs.PikaBodyGenerate22T2vGenerate22T2vPost(
+        initial_operation = await sync_op(
+            cls,
+            ApiEndpoint(path=PATH_TEXT_TO_VIDEO, method="POST"),
+            response_model=pika_defs.PikaGenerateResponse,
+            data=pika_defs.PikaBodyGenerate22T2vGenerate22T2vPost(
                 promptText=prompt_text,
                 negativePrompt=negative_prompt,
                 seed=seed,
@@ -202,10 +176,9 @@ class PikaTextToVideoNode(IO.ComfyNode):
                 duration=duration,
                 aspectRatio=aspect_ratio,
             ),
-            auth_kwargs=auth,
             content_type="application/x-www-form-urlencoded",
         )
-        return await execute_task(initial_operation, auth_kwargs=auth, node_id=cls.hidden.unique_id)
+        return await execute_task(initial_operation.video_id, cls)
 
 
 class PikaScenes(IO.ComfyNode):
@@ -309,24 +282,16 @@ class PikaScenes(IO.ComfyNode):
             duration=duration,
             aspectRatio=aspect_ratio,
         )
-        auth = {
-            "auth_token": cls.hidden.auth_token_comfy_org,
-            "comfy_api_key": cls.hidden.api_key_comfy_org,
-        }
-        initial_operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path=PATH_PIKASCENES,
-                method=HttpMethod.POST,
-                request_model=pika_defs.PikaBodyGenerate22C2vGenerate22PikascenesPost,
-                response_model=pika_defs.PikaGenerateResponse,
-            ),
-            request=pika_request_data,
+        initial_operation = await sync_op(
+            cls,
+            ApiEndpoint(path=PATH_PIKASCENES, method="POST"),
+            response_model=pika_defs.PikaGenerateResponse,
+            data=pika_request_data,
             files=pika_files,
             content_type="multipart/form-data",
-            auth_kwargs=auth,
         )
 
-        return await execute_task(initial_operation, auth_kwargs=auth, node_id=cls.hidden.unique_id)
+        return await execute_task(initial_operation.video_id, cls)
 
 
 class PikAdditionsNode(IO.ComfyNode):
@@ -383,24 +348,16 @@ class PikAdditionsNode(IO.ComfyNode):
             negativePrompt=negative_prompt,
             seed=seed,
         )
-        auth = {
-            "auth_token": cls.hidden.auth_token_comfy_org,
-            "comfy_api_key": cls.hidden.api_key_comfy_org,
-        }
-        initial_operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path=PATH_PIKADDITIONS,
-                method=HttpMethod.POST,
-                request_model=pika_defs.PikaBodyGeneratePikadditionsGeneratePikadditionsPost,
-                response_model=pika_defs.PikaGenerateResponse,
-            ),
-            request=pika_request_data,
+        initial_operation = await sync_op(
+            cls,
+            ApiEndpoint(path=PATH_PIKADDITIONS, method="POST"),
+            response_model=pika_defs.PikaGenerateResponse,
+            data=pika_request_data,
             files=pika_files,
             content_type="multipart/form-data",
-            auth_kwargs=auth,
         )
 
-        return await execute_task(initial_operation, auth_kwargs=auth, node_id=cls.hidden.unique_id)
+        return await execute_task(initial_operation.video_id, cls)
 
 
 class PikaSwapsNode(IO.ComfyNode):
@@ -472,23 +429,15 @@ class PikaSwapsNode(IO.ComfyNode):
             seed=seed,
             modifyRegionRoi=region_to_modify if region_to_modify else None,
         )
-        auth = {
-            "auth_token": cls.hidden.auth_token_comfy_org,
-            "comfy_api_key": cls.hidden.api_key_comfy_org,
-        }
-        initial_operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path=PATH_PIKASWAPS,
-                method=HttpMethod.POST,
-                request_model=pika_defs.PikaBodyGeneratePikaswapsGeneratePikaswapsPost,
-                response_model=pika_defs.PikaGenerateResponse,
-            ),
-            request=pika_request_data,
+        initial_operation = await sync_op(
+            cls,
+            ApiEndpoint(path=PATH_PIKASWAPS, method="POST"),
+            response_model=pika_defs.PikaGenerateResponse,
+            data=pika_request_data,
             files=pika_files,
             content_type="multipart/form-data",
-            auth_kwargs=auth,
         )
-        return await execute_task(initial_operation, auth_kwargs=auth, node_id=cls.hidden.unique_id)
+        return await execute_task(initial_operation.video_id, cls)
 
 
 class PikaffectsNode(IO.ComfyNode):
@@ -528,18 +477,11 @@ class PikaffectsNode(IO.ComfyNode):
         negative_prompt: str,
         seed: int,
     ) -> IO.NodeOutput:
-        auth = {
-            "auth_token": cls.hidden.auth_token_comfy_org,
-            "comfy_api_key": cls.hidden.api_key_comfy_org,
-        }
-        initial_operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path=PATH_PIKAFFECTS,
-                method=HttpMethod.POST,
-                request_model=pika_defs.PikaBodyGeneratePikaffectsGeneratePikaffectsPost,
-                response_model=pika_defs.PikaGenerateResponse,
-            ),
-            request=pika_defs.PikaBodyGeneratePikaffectsGeneratePikaffectsPost(
+        initial_operation = await sync_op(
+            cls,
+            ApiEndpoint(path=PATH_PIKAFFECTS, method="POST"),
+            response_model=pika_defs.PikaGenerateResponse,
+            data=pika_defs.PikaBodyGeneratePikaffectsGeneratePikaffectsPost(
                 pikaffect=pikaffect,
                 promptText=prompt_text,
                 negativePrompt=negative_prompt,
@@ -547,9 +489,8 @@ class PikaffectsNode(IO.ComfyNode):
             ),
             files={"image": ("image.png", tensor_to_bytesio(image), "image/png")},
             content_type="multipart/form-data",
-            auth_kwargs=auth,
         )
-        return await execute_task(initial_operation, auth_kwargs=auth, node_id=cls.hidden.unique_id)
+        return await execute_task(initial_operation.video_id, cls)
 
 
 class PikaStartEndFrameNode(IO.ComfyNode):
@@ -592,18 +533,11 @@ class PikaStartEndFrameNode(IO.ComfyNode):
             ("keyFrames", ("image_start.png", tensor_to_bytesio(image_start), "image/png")),
             ("keyFrames", ("image_end.png", tensor_to_bytesio(image_end), "image/png")),
         ]
-        auth = {
-            "auth_token": cls.hidden.auth_token_comfy_org,
-            "comfy_api_key": cls.hidden.api_key_comfy_org,
-        }
-        initial_operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path=PATH_PIKAFRAMES,
-                method=HttpMethod.POST,
-                request_model=pika_defs.PikaBodyGenerate22KeyframeGenerate22PikaframesPost,
-                response_model=pika_defs.PikaGenerateResponse,
-            ),
-            request=pika_defs.PikaBodyGenerate22KeyframeGenerate22PikaframesPost(
+        initial_operation = await sync_op(
+            cls,
+            ApiEndpoint(path=PATH_PIKAFRAMES, method="POST"),
+            response_model=pika_defs.PikaGenerateResponse,
+            data=pika_defs.PikaBodyGenerate22KeyframeGenerate22PikaframesPost(
                 promptText=prompt_text,
                 negativePrompt=negative_prompt,
                 seed=seed,
@@ -612,9 +546,8 @@ class PikaStartEndFrameNode(IO.ComfyNode):
             ),
             files=pika_files,
             content_type="multipart/form-data",
-            auth_kwargs=auth,
         )
-        return await execute_task(initial_operation, auth_kwargs=auth, node_id=cls.hidden.unique_id)
+        return await execute_task(initial_operation.video_id, cls)
 
 
 class PikaApiNodesExtension(ComfyExtension):

comfy_api_nodes/nodes_pixverse.py

@@ -1,7 +1,6 @@
-from inspect import cleandoc
-from typing import Optional
+import torch
 from typing_extensions import override
-from io import BytesIO
+from comfy_api.latest import IO, ComfyExtension
 from comfy_api_nodes.apis.pixverse_api import (
     PixverseTextVideoRequest,
     PixverseImageVideoRequest,
@@ -17,53 +16,30 @@ from comfy_api_nodes.apis.pixverse_api import (
     PixverseIO,
     pixverse_templates,
 )
-from comfy_api_nodes.apis.client import (
+from comfy_api_nodes.util import (
     ApiEndpoint,
-    HttpMethod,
-    SynchronousOperation,
-    PollingOperation,
-    EmptyRequest,
+    download_url_to_video_output,
+    poll_op,
+    sync_op,
+    tensor_to_bytesio,
+    validate_string,
 )
-from comfy_api_nodes.util import validate_string, tensor_to_bytesio
-from comfy_api.input_impl import VideoFromFile
-from comfy_api.latest import ComfyExtension, IO
-
-import torch
-import aiohttp
-
 
 AVERAGE_DURATION_T2V = 32
 AVERAGE_DURATION_I2V = 30
 AVERAGE_DURATION_T2T = 52
 
 
-def get_video_url_from_response(
-    response: PixverseGenerationStatusResponse,
-) -> Optional[str]:
-    if response.Resp is None or response.Resp.url is None:
-        return None
-    return str(response.Resp.url)
-
-
-async def upload_image_to_pixverse(image: torch.Tensor, auth_kwargs=None):
-    # first, upload image to Pixverse and get image id to use in actual generation call
-    operation = SynchronousOperation(
-        endpoint=ApiEndpoint(
-            path="/proxy/pixverse/image/upload",
-            method=HttpMethod.POST,
-            request_model=EmptyRequest,
-            response_model=PixverseImageUploadResponse,
-        ),
-        request=EmptyRequest(),
+async def upload_image_to_pixverse(cls: type[IO.ComfyNode], image: torch.Tensor):
+    response_upload = await sync_op(
+        cls,
+        ApiEndpoint(path="/proxy/pixverse/image/upload", method="POST"),
+        response_model=PixverseImageUploadResponse,
         files={"image": tensor_to_bytesio(image)},
         content_type="multipart/form-data",
-        auth_kwargs=auth_kwargs,
     )
-    response_upload: PixverseImageUploadResponse = await operation.execute()
-
     if response_upload.Resp is None:
         raise Exception(f"PixVerse image upload request failed: '{response_upload.ErrMsg}'")
-
     return response_upload.Resp.img_id
 
 
@@ -93,17 +69,13 @@ class PixverseTemplateNode(IO.ComfyNode):
 
 
 class PixverseTextToVideoNode(IO.ComfyNode):
-    """
-    Generates videos based on prompt and output_size.
-    """
-
     @classmethod
     def define_schema(cls) -> IO.Schema:
         return IO.Schema(
             node_id="PixverseTextToVideoNode",
             display_name="PixVerse Text to Video",
             category="api node/video/PixVerse",
-            description=cleandoc(cls.__doc__ or ""),
+            description="Generates videos based on prompt and output_size.",
             inputs=[
                 IO.String.Input(
                     "prompt",
@@ -170,7 +142,7 @@ class PixverseTextToVideoNode(IO.ComfyNode):
         negative_prompt: str = None,
         pixverse_template: int = None,
     ) -> IO.NodeOutput:
-        validate_string(prompt, strip_whitespace=False)
+        validate_string(prompt, strip_whitespace=False, min_length=1)
         # 1080p is limited to 5 seconds duration
         # only normal motion_mode supported for 1080p or for non-5 second duration
         if quality == PixverseQuality.res_1080p:
@@ -179,18 +151,11 @@ class PixverseTextToVideoNode(IO.ComfyNode):
         elif duration_seconds != PixverseDuration.dur_5:
             motion_mode = PixverseMotionMode.normal
 
-        auth = {
-            "auth_token": cls.hidden.auth_token_comfy_org,
-            "comfy_api_key": cls.hidden.api_key_comfy_org,
-        }
-        operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path="/proxy/pixverse/video/text/generate",
-                method=HttpMethod.POST,
-                request_model=PixverseTextVideoRequest,
-                response_model=PixverseVideoResponse,
-            ),
-            request=PixverseTextVideoRequest(
+        response_api = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/pixverse/video/text/generate", method="POST"),
+            response_model=PixverseVideoResponse,
+            data=PixverseTextVideoRequest(
                 prompt=prompt,
                 aspect_ratio=aspect_ratio,
                 quality=quality,
@@ -200,20 +165,14 @@ class PixverseTextToVideoNode(IO.ComfyNode):
                 template_id=pixverse_template,
                 seed=seed,
             ),
-            auth_kwargs=auth,
         )
-        response_api = await operation.execute()
-
         if response_api.Resp is None:
             raise Exception(f"PixVerse request failed: '{response_api.ErrMsg}'")
 
-        operation = PollingOperation(
-            poll_endpoint=ApiEndpoint(
-                path=f"/proxy/pixverse/video/result/{response_api.Resp.video_id}",
-                method=HttpMethod.GET,
-                request_model=EmptyRequest,
-                response_model=PixverseGenerationStatusResponse,
-            ),
+        response_poll = await poll_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/pixverse/video/result/{response_api.Resp.video_id}"),
+            response_model=PixverseGenerationStatusResponse,
             completed_statuses=[PixverseStatus.successful],
             failed_statuses=[
                 PixverseStatus.contents_moderation,
@@ -221,30 +180,19 @@ class PixverseTextToVideoNode(IO.ComfyNode):
                 PixverseStatus.deleted,
             ],
             status_extractor=lambda x: x.Resp.status,
-            auth_kwargs=auth,
-            node_id=cls.hidden.unique_id,
-            result_url_extractor=get_video_url_from_response,
             estimated_duration=AVERAGE_DURATION_T2V,
         )
-        response_poll = await operation.execute()
-
-        async with aiohttp.ClientSession() as session:
-            async with session.get(response_poll.Resp.url) as vid_response:
-                return IO.NodeOutput(VideoFromFile(BytesIO(await vid_response.content.read())))
+        return IO.NodeOutput(await download_url_to_video_output(response_poll.Resp.url))
 
 
 class PixverseImageToVideoNode(IO.ComfyNode):
-    """
-    Generates videos based on prompt and output_size.
-    """
-
     @classmethod
     def define_schema(cls) -> IO.Schema:
         return IO.Schema(
             node_id="PixverseImageToVideoNode",
             display_name="PixVerse Image to Video",
             category="api node/video/PixVerse",
-            description=cleandoc(cls.__doc__ or ""),
+            description="Generates videos based on prompt and output_size.",
             inputs=[
                 IO.Image.Input("image"),
                 IO.String.Input(
@@ -309,11 +257,7 @@ class PixverseImageToVideoNode(IO.ComfyNode):
         pixverse_template: int = None,
     ) -> IO.NodeOutput:
         validate_string(prompt, strip_whitespace=False)
-        auth = {
-            "auth_token": cls.hidden.auth_token_comfy_org,
-            "comfy_api_key": cls.hidden.api_key_comfy_org,
-        }
-        img_id = await upload_image_to_pixverse(image, auth_kwargs=auth)
+        img_id = await upload_image_to_pixverse(cls, image)
 
         # 1080p is limited to 5 seconds duration
         # only normal motion_mode supported for 1080p or for non-5 second duration
@@ -323,14 +267,11 @@ class PixverseImageToVideoNode(IO.ComfyNode):
         elif duration_seconds != PixverseDuration.dur_5:
             motion_mode = PixverseMotionMode.normal
 
-        operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path="/proxy/pixverse/video/img/generate",
-                method=HttpMethod.POST,
-                request_model=PixverseImageVideoRequest,
-                response_model=PixverseVideoResponse,
-            ),
-            request=PixverseImageVideoRequest(
+        response_api = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/pixverse/video/img/generate", method="POST"),
+            response_model=PixverseVideoResponse,
+            data=PixverseImageVideoRequest(
                 img_id=img_id,
                 prompt=prompt,
                 quality=quality,
@@ -340,20 +281,15 @@ class PixverseImageToVideoNode(IO.ComfyNode):
                 template_id=pixverse_template,
                 seed=seed,
             ),
-            auth_kwargs=auth,
         )
-        response_api = await operation.execute()
 
         if response_api.Resp is None:
             raise Exception(f"PixVerse request failed: '{response_api.ErrMsg}'")
 
-        operation = PollingOperation(
-            poll_endpoint=ApiEndpoint(
-                path=f"/proxy/pixverse/video/result/{response_api.Resp.video_id}",
-                method=HttpMethod.GET,
-                request_model=EmptyRequest,
-                response_model=PixverseGenerationStatusResponse,
-            ),
+        response_poll = await poll_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/pixverse/video/result/{response_api.Resp.video_id}"),
+            response_model=PixverseGenerationStatusResponse,
             completed_statuses=[PixverseStatus.successful],
             failed_statuses=[
                 PixverseStatus.contents_moderation,
@@ -361,30 +297,19 @@ class PixverseImageToVideoNode(IO.ComfyNode):
                 PixverseStatus.deleted,
             ],
             status_extractor=lambda x: x.Resp.status,
-            auth_kwargs=auth,
-            node_id=cls.hidden.unique_id,
-            result_url_extractor=get_video_url_from_response,
             estimated_duration=AVERAGE_DURATION_I2V,
         )
-        response_poll = await operation.execute()
-
-        async with aiohttp.ClientSession() as session:
-            async with session.get(response_poll.Resp.url) as vid_response:
-                return IO.NodeOutput(VideoFromFile(BytesIO(await vid_response.content.read())))
+        return IO.NodeOutput(await download_url_to_video_output(response_poll.Resp.url))
 
 
 class PixverseTransitionVideoNode(IO.ComfyNode):
-    """
-    Generates videos based on prompt and output_size.
-    """
-
     @classmethod
     def define_schema(cls) -> IO.Schema:
         return IO.Schema(
             node_id="PixverseTransitionVideoNode",
             display_name="PixVerse Transition Video",
             category="api node/video/PixVerse",
-            description=cleandoc(cls.__doc__ or ""),
+            description="Generates videos based on prompt and output_size.",
             inputs=[
                 IO.Image.Input("first_frame"),
                 IO.Image.Input("last_frame"),
@@ -445,12 +370,8 @@ class PixverseTransitionVideoNode(IO.ComfyNode):
         negative_prompt: str = None,
     ) -> IO.NodeOutput:
         validate_string(prompt, strip_whitespace=False)
-        auth = {
-            "auth_token": cls.hidden.auth_token_comfy_org,
-            "comfy_api_key": cls.hidden.api_key_comfy_org,
-        }
-        first_frame_id = await upload_image_to_pixverse(first_frame, auth_kwargs=auth)
-        last_frame_id = await upload_image_to_pixverse(last_frame, auth_kwargs=auth)
+        first_frame_id = await upload_image_to_pixverse(cls, first_frame)
+        last_frame_id = await upload_image_to_pixverse(cls, last_frame)
 
         # 1080p is limited to 5 seconds duration
         # only normal motion_mode supported for 1080p or for non-5 second duration
@@ -460,14 +381,11 @@ class PixverseTransitionVideoNode(IO.ComfyNode):
         elif duration_seconds != PixverseDuration.dur_5:
             motion_mode = PixverseMotionMode.normal
 
-        operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path="/proxy/pixverse/video/transition/generate",
-                method=HttpMethod.POST,
-                request_model=PixverseTransitionVideoRequest,
-                response_model=PixverseVideoResponse,
-            ),
-            request=PixverseTransitionVideoRequest(
+        response_api = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/pixverse/video/transition/generate", method="POST"),
+            response_model=PixverseVideoResponse,
+            data=PixverseTransitionVideoRequest(
                 first_frame_img=first_frame_id,
                 last_frame_img=last_frame_id,
                 prompt=prompt,
@@ -477,20 +395,15 @@ class PixverseTransitionVideoNode(IO.ComfyNode):
                 negative_prompt=negative_prompt if negative_prompt else None,
                 seed=seed,
             ),
-            auth_kwargs=auth,
         )
-        response_api = await operation.execute()
 
         if response_api.Resp is None:
             raise Exception(f"PixVerse request failed: '{response_api.ErrMsg}'")
 
-        operation = PollingOperation(
-            poll_endpoint=ApiEndpoint(
-                path=f"/proxy/pixverse/video/result/{response_api.Resp.video_id}",
-                method=HttpMethod.GET,
-                request_model=EmptyRequest,
-                response_model=PixverseGenerationStatusResponse,
-            ),
+        response_poll = await poll_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/pixverse/video/result/{response_api.Resp.video_id}"),
+            response_model=PixverseGenerationStatusResponse,
             completed_statuses=[PixverseStatus.successful],
             failed_statuses=[
                 PixverseStatus.contents_moderation,
@@ -498,16 +411,9 @@ class PixverseTransitionVideoNode(IO.ComfyNode):
                 PixverseStatus.deleted,
             ],
             status_extractor=lambda x: x.Resp.status,
-            auth_kwargs=auth,
-            node_id=cls.hidden.unique_id,
-            result_url_extractor=get_video_url_from_response,
             estimated_duration=AVERAGE_DURATION_T2V,
         )
-        response_poll = await operation.execute()
-
-        async with aiohttp.ClientSession() as session:
-            async with session.get(response_poll.Resp.url) as vid_response:
-                return IO.NodeOutput(VideoFromFile(BytesIO(await vid_response.content.read())))
+        return IO.NodeOutput(await download_url_to_video_output(response_poll.Resp.url))
 
 
 class PixVerseExtension(ComfyExtension):

comfy_api_nodes/nodes_recraft.py

@@ -8,9 +8,6 @@ from typing_extensions import override
 
 from comfy.utils import ProgressBar
 from comfy_api.latest import IO, ComfyExtension
-from comfy_api_nodes.apinode_utils import (
-    resize_mask_to_image,
-)
 from comfy_api_nodes.apis.recraft_api import (
     RecraftColor,
     RecraftColorChain,
@@ -28,6 +25,7 @@ from comfy_api_nodes.util import (
     ApiEndpoint,
     bytesio_to_image_tensor,
     download_url_as_bytesio,
+    resize_mask_to_image,
     sync_op,
     tensor_to_bytesio,
     validate_string,

comfy_api_nodes/nodes_rodin.py

@@ -225,21 +225,20 @@ async def get_rodin_download_list(uuid, auth_kwargs: Optional[dict[str, str]] =
 
 
 async def download_files(url_list, task_uuid):
-    save_path = os.path.join(comfy_paths.get_output_directory(), f"Rodin3D_{task_uuid}")
+    result_folder_name = f"Rodin3D_{task_uuid}"
+    save_path = os.path.join(comfy_paths.get_output_directory(), result_folder_name)
     os.makedirs(save_path, exist_ok=True)
     model_file_path = None
     async with aiohttp.ClientSession() as session:
         for i in url_list.list:
-            url = i.url
-            file_name = i.name
-            file_path = os.path.join(save_path, file_name)
+            file_path = os.path.join(save_path, i.name)
             if file_path.endswith(".glb"):
-                model_file_path = file_path
+                model_file_path = os.path.join(result_folder_name, i.name)
             logging.info("[ Rodin3D API - download_files ] Downloading file: %s", file_path)
             max_retries = 5
             for attempt in range(max_retries):
                 try:
-                    async with session.get(url) as resp:
+                    async with session.get(i.url) as resp:
                         resp.raise_for_status()
                         with open(file_path, "wb") as f:
                             async for chunk in resp.content.iter_chunked(32 * 1024):

comfy_api_nodes/nodes_runway.py

@@ -200,7 +200,7 @@ class RunwayImageToVideoNodeGen3a(IO.ComfyNode):
     ) -> IO.NodeOutput:
         validate_string(prompt, min_length=1)
         validate_image_dimensions(start_frame, max_width=7999, max_height=7999)
-        validate_image_aspect_ratio(start_frame, min_aspect_ratio=0.5, max_aspect_ratio=2.0)
+        validate_image_aspect_ratio(start_frame, (1, 2), (2, 1))
 
         download_urls = await upload_images_to_comfyapi(
             cls,
@@ -290,7 +290,7 @@ class RunwayImageToVideoNodeGen4(IO.ComfyNode):
     ) -> IO.NodeOutput:
         validate_string(prompt, min_length=1)
         validate_image_dimensions(start_frame, max_width=7999, max_height=7999)
-        validate_image_aspect_ratio(start_frame, min_aspect_ratio=0.5, max_aspect_ratio=2.0)
+        validate_image_aspect_ratio(start_frame, (1, 2), (2, 1))
 
         download_urls = await upload_images_to_comfyapi(
             cls,
@@ -390,8 +390,8 @@ class RunwayFirstLastFrameNode(IO.ComfyNode):
         validate_string(prompt, min_length=1)
         validate_image_dimensions(start_frame, max_width=7999, max_height=7999)
         validate_image_dimensions(end_frame, max_width=7999, max_height=7999)
-        validate_image_aspect_ratio(start_frame, min_aspect_ratio=0.5, max_aspect_ratio=2.0)
-        validate_image_aspect_ratio(end_frame, min_aspect_ratio=0.5, max_aspect_ratio=2.0)
+        validate_image_aspect_ratio(start_frame, (1, 2), (2, 1))
+        validate_image_aspect_ratio(end_frame, (1, 2), (2, 1))
 
         stacked_input_images = image_tensor_pair_to_batch(start_frame, end_frame)
         download_urls = await upload_images_to_comfyapi(
@@ -475,7 +475,7 @@ class RunwayTextToImageNode(IO.ComfyNode):
         reference_images = None
         if reference_image is not None:
             validate_image_dimensions(reference_image, max_width=7999, max_height=7999)
-            validate_image_aspect_ratio(reference_image, min_aspect_ratio=0.5, max_aspect_ratio=2.0)
+            validate_image_aspect_ratio(reference_image, (1, 2), (2, 1))
             download_urls = await upload_images_to_comfyapi(
                 cls,
                 reference_image,

comfy_api_nodes/nodes_stability.py

@@ -20,13 +20,6 @@ from comfy_api_nodes.apis.stability_api import (
     StabilityAudioInpaintRequest,
     StabilityAudioResponse,
 )
-from comfy_api_nodes.apis.client import (
-    ApiEndpoint,
-    HttpMethod,
-    SynchronousOperation,
-    PollingOperation,
-    EmptyRequest,
-)
 from comfy_api_nodes.util import (
     validate_audio_duration,
     validate_string,
@@ -34,6 +27,9 @@ from comfy_api_nodes.util import (
     bytesio_to_image_tensor,
     tensor_to_bytesio,
     audio_bytes_to_audio_input,
+    sync_op,
+    poll_op,
+    ApiEndpoint,
 )
 
 import torch
@@ -161,19 +157,11 @@ class StabilityStableImageUltraNode(IO.ComfyNode):
             "image": image_binary
         }
 
-        auth = {
-            "auth_token": cls.hidden.auth_token_comfy_org,
-            "comfy_api_key": cls.hidden.api_key_comfy_org,
-        }
-
-        operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path="/proxy/stability/v2beta/stable-image/generate/ultra",
-                method=HttpMethod.POST,
-                request_model=StabilityStableUltraRequest,
-                response_model=StabilityStableUltraResponse,
-            ),
-            request=StabilityStableUltraRequest(
+        response_api = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/stability/v2beta/stable-image/generate/ultra", method="POST"),
+            response_model=StabilityStableUltraResponse,
+            data=StabilityStableUltraRequest(
                 prompt=prompt,
                 negative_prompt=negative_prompt,
                 aspect_ratio=aspect_ratio,
@@ -183,9 +171,7 @@ class StabilityStableImageUltraNode(IO.ComfyNode):
             ),
             files=files,
             content_type="multipart/form-data",
-            auth_kwargs=auth,
         )
-        response_api = await operation.execute()
 
         if response_api.finish_reason != "SUCCESS":
             raise Exception(f"Stable Image Ultra generation failed: {response_api.finish_reason}.")
@@ -313,19 +299,11 @@ class StabilityStableImageSD_3_5Node(IO.ComfyNode):
             "image": image_binary
         }
 
-        auth = {
-            "auth_token": cls.hidden.auth_token_comfy_org,
-            "comfy_api_key": cls.hidden.api_key_comfy_org,
-        }
-
-        operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path="/proxy/stability/v2beta/stable-image/generate/sd3",
-                method=HttpMethod.POST,
-                request_model=StabilityStable3_5Request,
-                response_model=StabilityStableUltraResponse,
-            ),
-            request=StabilityStable3_5Request(
+        response_api = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/stability/v2beta/stable-image/generate/sd3", method="POST"),
+            response_model=StabilityStableUltraResponse,
+            data=StabilityStable3_5Request(
                 prompt=prompt,
                 negative_prompt=negative_prompt,
                 aspect_ratio=aspect_ratio,
@@ -338,9 +316,7 @@ class StabilityStableImageSD_3_5Node(IO.ComfyNode):
             ),
             files=files,
             content_type="multipart/form-data",
-            auth_kwargs=auth,
         )
-        response_api = await operation.execute()
 
         if response_api.finish_reason != "SUCCESS":
             raise Exception(f"Stable Diffusion 3.5 Image generation failed: {response_api.finish_reason}.")
@@ -427,19 +403,11 @@ class StabilityUpscaleConservativeNode(IO.ComfyNode):
             "image": image_binary
         }
 
-        auth = {
-            "auth_token": cls.hidden.auth_token_comfy_org,
-            "comfy_api_key": cls.hidden.api_key_comfy_org,
-        }
-
-        operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path="/proxy/stability/v2beta/stable-image/upscale/conservative",
-                method=HttpMethod.POST,
-                request_model=StabilityUpscaleConservativeRequest,
-                response_model=StabilityStableUltraResponse,
-            ),
-            request=StabilityUpscaleConservativeRequest(
+        response_api = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/stability/v2beta/stable-image/upscale/conservative", method="POST"),
+            response_model=StabilityStableUltraResponse,
+            data=StabilityUpscaleConservativeRequest(
                 prompt=prompt,
                 negative_prompt=negative_prompt,
                 creativity=round(creativity,2),
@@ -447,9 +415,7 @@ class StabilityUpscaleConservativeNode(IO.ComfyNode):
             ),
             files=files,
             content_type="multipart/form-data",
-            auth_kwargs=auth,
         )
-        response_api = await operation.execute()
 
         if response_api.finish_reason != "SUCCESS":
             raise Exception(f"Stability Upscale Conservative generation failed: {response_api.finish_reason}.")
@@ -544,19 +510,11 @@ class StabilityUpscaleCreativeNode(IO.ComfyNode):
             "image": image_binary
         }
 
-        auth = {
-            "auth_token": cls.hidden.auth_token_comfy_org,
-            "comfy_api_key": cls.hidden.api_key_comfy_org,
-        }
-
-        operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path="/proxy/stability/v2beta/stable-image/upscale/creative",
-                method=HttpMethod.POST,
-                request_model=StabilityUpscaleCreativeRequest,
-                response_model=StabilityAsyncResponse,
-            ),
-            request=StabilityUpscaleCreativeRequest(
+        response_api = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/stability/v2beta/stable-image/upscale/creative", method="POST"),
+            response_model=StabilityAsyncResponse,
+            data=StabilityUpscaleCreativeRequest(
                 prompt=prompt,
                 negative_prompt=negative_prompt,
                 creativity=round(creativity,2),
@@ -565,25 +523,15 @@ class StabilityUpscaleCreativeNode(IO.ComfyNode):
             ),
             files=files,
             content_type="multipart/form-data",
-            auth_kwargs=auth,
         )
-        response_api = await operation.execute()
 
-        operation = PollingOperation(
-            poll_endpoint=ApiEndpoint(
-                path=f"/proxy/stability/v2beta/results/{response_api.id}",
-                method=HttpMethod.GET,
-                request_model=EmptyRequest,
-                response_model=StabilityResultsGetResponse,
-            ),
+        response_poll = await poll_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/stability/v2beta/results/{response_api.id}"),
+            response_model=StabilityResultsGetResponse,
             poll_interval=3,
-            completed_statuses=[StabilityPollStatus.finished],
-            failed_statuses=[StabilityPollStatus.failed],
             status_extractor=lambda x: get_async_dummy_status(x),
-            auth_kwargs=auth,
-            node_id=cls.hidden.unique_id,
         )
-        response_poll: StabilityResultsGetResponse = await operation.execute()
 
         if response_poll.finish_reason != "SUCCESS":
             raise Exception(f"Stability Upscale Creative generation failed: {response_poll.finish_reason}.")
@@ -628,24 +576,13 @@ class StabilityUpscaleFastNode(IO.ComfyNode):
             "image": image_binary
         }
 
-        auth = {
-            "auth_token": cls.hidden.auth_token_comfy_org,
-            "comfy_api_key": cls.hidden.api_key_comfy_org,
-        }
-
-        operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path="/proxy/stability/v2beta/stable-image/upscale/fast",
-                method=HttpMethod.POST,
-                request_model=EmptyRequest,
-                response_model=StabilityStableUltraResponse,
-            ),
-            request=EmptyRequest(),
+        response_api = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/stability/v2beta/stable-image/upscale/fast", method="POST"),
+            response_model=StabilityStableUltraResponse,
             files=files,
             content_type="multipart/form-data",
-            auth_kwargs=auth,
         )
-        response_api = await operation.execute()
 
         if response_api.finish_reason != "SUCCESS":
             raise Exception(f"Stability Upscale Fast failed: {response_api.finish_reason}.")
@@ -717,21 +654,13 @@ class StabilityTextToAudio(IO.ComfyNode):
     async def execute(cls, model: str, prompt: str, duration: int, seed: int, steps: int) -> IO.NodeOutput:
         validate_string(prompt, max_length=10000)
         payload = StabilityTextToAudioRequest(prompt=prompt, model=model, duration=duration, seed=seed, steps=steps)
-        operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path="/proxy/stability/v2beta/audio/stable-audio-2/text-to-audio",
-                method=HttpMethod.POST,
-                request_model=StabilityTextToAudioRequest,
-                response_model=StabilityAudioResponse,
-            ),
-            request=payload,
+        response_api = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/stability/v2beta/audio/stable-audio-2/text-to-audio", method="POST"),
+            response_model=StabilityAudioResponse,
+            data=payload,
             content_type="multipart/form-data",
-            auth_kwargs= {
-                "auth_token": cls.hidden.auth_token_comfy_org,
-                "comfy_api_key": cls.hidden.api_key_comfy_org,
-            },
         )
-        response_api = await operation.execute()
         if not response_api.audio:
             raise ValueError("No audio file was received in response.")
         return IO.NodeOutput(audio_bytes_to_audio_input(base64.b64decode(response_api.audio)))
@@ -814,22 +743,14 @@ class StabilityAudioToAudio(IO.ComfyNode):
         payload = StabilityAudioToAudioRequest(
             prompt=prompt, model=model, duration=duration, seed=seed, steps=steps, strength=strength
         )
-        operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path="/proxy/stability/v2beta/audio/stable-audio-2/audio-to-audio",
-                method=HttpMethod.POST,
-                request_model=StabilityAudioToAudioRequest,
-                response_model=StabilityAudioResponse,
-            ),
-            request=payload,
+        response_api = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/stability/v2beta/audio/stable-audio-2/audio-to-audio", method="POST"),
+            response_model=StabilityAudioResponse,
+            data=payload,
             content_type="multipart/form-data",
             files={"audio": audio_input_to_mp3(audio)},
-            auth_kwargs= {
-                "auth_token": cls.hidden.auth_token_comfy_org,
-                "comfy_api_key": cls.hidden.api_key_comfy_org,
-            },
         )
-        response_api = await operation.execute()
         if not response_api.audio:
             raise ValueError("No audio file was received in response.")
         return IO.NodeOutput(audio_bytes_to_audio_input(base64.b64decode(response_api.audio)))
@@ -935,22 +856,14 @@ class StabilityAudioInpaint(IO.ComfyNode):
             mask_start=mask_start,
             mask_end=mask_end,
         )
-        operation = SynchronousOperation(
-            endpoint=ApiEndpoint(
-                path="/proxy/stability/v2beta/audio/stable-audio-2/inpaint",
-                method=HttpMethod.POST,
-                request_model=StabilityAudioInpaintRequest,
-                response_model=StabilityAudioResponse,
-            ),
-            request=payload,
+        response_api = await sync_op(
+            cls,
+            endpoint=ApiEndpoint(path="/proxy/stability/v2beta/audio/stable-audio-2/inpaint", method="POST"),
+            response_model=StabilityAudioResponse,
+            data=payload,
             content_type="multipart/form-data",
             files={"audio": audio_input_to_mp3(audio)},
-            auth_kwargs={
-                "auth_token": cls.hidden.auth_token_comfy_org,
-                "comfy_api_key": cls.hidden.api_key_comfy_org,
-            },
         )
-        response_api = await operation.execute()
         if not response_api.audio:
             raise ValueError("No audio file was received in response.")
         return IO.NodeOutput(audio_bytes_to_audio_input(base64.b64decode(response_api.audio)))

comfy_api_nodes/nodes_vidu.py

@@ -14,9 +14,9 @@ from comfy_api_nodes.util import (
     poll_op,
     sync_op,
     upload_images_to_comfyapi,
-    validate_aspect_ratio_closeness,
-    validate_image_aspect_ratio_range,
+    validate_image_aspect_ratio,
     validate_image_dimensions,
+    validate_images_aspect_ratio_closeness,
 )
 
 VIDU_TEXT_TO_VIDEO = "/proxy/vidu/text2video"
@@ -114,7 +114,7 @@ async def execute_task(
         cls,
         ApiEndpoint(path=VIDU_GET_GENERATION_STATUS % response.task_id),
         response_model=TaskStatusResponse,
-        status_extractor=lambda r: r.state.value,
+        status_extractor=lambda r: r.state,
         estimated_duration=estimated_duration,
     )
 
@@ -307,7 +307,7 @@ class ViduImageToVideoNode(IO.ComfyNode):
     ) -> IO.NodeOutput:
         if get_number_of_images(image) > 1:
             raise ValueError("Only one input image is allowed.")
-        validate_image_aspect_ratio_range(image, (1, 4), (4, 1))
+        validate_image_aspect_ratio(image, (1, 4), (4, 1))
         payload = TaskCreationRequest(
             model_name=model,
             prompt=prompt,
@@ -423,7 +423,7 @@ class ViduReferenceVideoNode(IO.ComfyNode):
         if a > 7:
             raise ValueError("Too many images, maximum allowed is 7.")
         for image in images:
-            validate_image_aspect_ratio_range(image, (1, 4), (4, 1))
+            validate_image_aspect_ratio(image, (1, 4), (4, 1))
             validate_image_dimensions(image, min_width=128, min_height=128)
         payload = TaskCreationRequest(
             model_name=model,
@@ -533,7 +533,7 @@ class ViduStartEndToVideoNode(IO.ComfyNode):
         resolution: str,
         movement_amplitude: str,
     ) -> IO.NodeOutput:
-        validate_aspect_ratio_closeness(first_frame, end_frame, min_rel=0.8, max_rel=1.25, strict=False)
+        validate_images_aspect_ratio_closeness(first_frame, end_frame, min_rel=0.8, max_rel=1.25, strict=False)
         payload = TaskCreationRequest(
             model_name=model,
             prompt=prompt,

comfy_api_nodes/util/__init__.py

@@ -14,9 +14,12 @@ from .conversions import (
     downscale_image_tensor,
     image_tensor_pair_to_batch,
     pil_to_bytesio,
+    resize_mask_to_image,
     tensor_to_base64_string,
     tensor_to_bytesio,
     tensor_to_pil,
+    text_filepath_to_base64_string,
+    text_filepath_to_data_uri,
     trim_video,
     video_to_base64_string,
 )
@@ -34,12 +37,12 @@ from .upload_helpers import (
 )
 from .validation_utils import (
     get_number_of_images,
-    validate_aspect_ratio_closeness,
+    validate_aspect_ratio_string,
     validate_audio_duration,
     validate_container_format_is_mp4,
     validate_image_aspect_ratio,
-    validate_image_aspect_ratio_range,
     validate_image_dimensions,
+    validate_images_aspect_ratio_closeness,
     validate_string,
     validate_video_dimensions,
     validate_video_duration,
@@ -70,19 +73,22 @@ __all__ = [
     "downscale_image_tensor",
     "image_tensor_pair_to_batch",
     "pil_to_bytesio",
+    "resize_mask_to_image",
     "tensor_to_base64_string",
     "tensor_to_bytesio",
     "tensor_to_pil",
+    "text_filepath_to_base64_string",
+    "text_filepath_to_data_uri",
     "trim_video",
     "video_to_base64_string",
     # Validation utilities
     "get_number_of_images",
-    "validate_aspect_ratio_closeness",
+    "validate_aspect_ratio_string",
     "validate_audio_duration",
     "validate_container_format_is_mp4",
     "validate_image_aspect_ratio",
-    "validate_image_aspect_ratio_range",
     "validate_image_dimensions",
+    "validate_images_aspect_ratio_closeness",
     "validate_string",
     "validate_video_dimensions",
     "validate_video_duration",

comfy_api_nodes/util/client.py

@@ -77,8 +77,8 @@ class _PollUIState:
 
 
 _RETRY_STATUS = {408, 429, 500, 502, 503, 504}
-COMPLETED_STATUSES = ["succeeded", "succeed", "success", "completed"]
-FAILED_STATUSES = ["cancelled", "canceled", "failed", "error"]
+COMPLETED_STATUSES = ["succeeded", "succeed", "success", "completed", "finished"]
+FAILED_STATUSES = ["cancelled", "canceled", "fail", "failed", "error"]
 QUEUED_STATUSES = ["created", "queued", "queueing", "submitted"]
 
 
@@ -589,7 +589,7 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool):
         operation_id = _generate_operation_id(method, cfg.endpoint.path, attempt)
         logging.debug("[DEBUG] HTTP %s %s (attempt %d)", method, url, attempt)
 
-        payload_headers = {"Accept": "*/*"}
+        payload_headers = {"Accept": "*/*"} if expect_binary else {"Accept": "application/json"}
         if not parsed_url.scheme and not parsed_url.netloc:  # is URL relative?
             payload_headers.update(get_auth_header(cfg.node_cls))
         if cfg.endpoint.headers:

comfy_api_nodes/util/conversions.py

@@ -1,6 +1,7 @@
 import base64
 import logging
 import math
+import mimetypes
 import uuid
 from io import BytesIO
 from typing import Optional
@@ -12,7 +13,7 @@ from PIL import Image
 
 from comfy.utils import common_upscale
 from comfy_api.latest import Input, InputImpl
-from comfy_api.util import VideoContainer, VideoCodec
+from comfy_api.util import VideoCodec, VideoContainer
 
 from ._helpers import mimetype_to_extension
 
@@ -430,3 +431,40 @@ def audio_bytes_to_audio_input(audio_bytes: bytes) -> dict:
     wav = torch.cat(frames, dim=1)  # [C, T]
     wav = _f32_pcm(wav)
     return {"waveform": wav.unsqueeze(0).contiguous(), "sample_rate": out_sr}
+
+
+def resize_mask_to_image(
+    mask: torch.Tensor,
+    image: torch.Tensor,
+    upscale_method="nearest-exact",
+    crop="disabled",
+    allow_gradient=True,
+    add_channel_dim=False,
+):
+    """Resize mask to be the same dimensions as an image, while maintaining proper format for API calls."""
+    _, height, width, _ = image.shape
+    mask = mask.unsqueeze(-1)
+    mask = mask.movedim(-1, 1)
+    mask = common_upscale(mask, width=width, height=height, upscale_method=upscale_method, crop=crop)
+    mask = mask.movedim(1, -1)
+    if not add_channel_dim:
+        mask = mask.squeeze(-1)
+    if not allow_gradient:
+        mask = (mask > 0.5).float()
+    return mask
+
+
+def text_filepath_to_base64_string(filepath: str) -> str:
+    """Converts a text file to a base64 string."""
+    with open(filepath, "rb") as f:
+        file_content = f.read()
+    return base64.b64encode(file_content).decode("utf-8")
+
+
+def text_filepath_to_data_uri(filepath: str) -> str:
+    """Converts a text file to a data URI."""
+    base64_string = text_filepath_to_base64_string(filepath)
+    mime_type, _ = mimetypes.guess_type(filepath)
+    if mime_type is None:
+        mime_type = "application/octet-stream"
+    return f"data:{mime_type};base64,{base64_string}"

comfy_api_nodes/util/download_helpers.py

@@ -232,11 +232,12 @@ async def download_url_to_video_output(
     video_url: str,
     *,
     timeout: float = None,
+    max_retries: int = 5,
     cls: type[COMFY_IO.ComfyNode] = None,
 ) -> VideoFromFile:
     """Downloads a video from a URL and returns a `VIDEO` output."""
     result = BytesIO()
-    await download_url_to_bytesio(video_url, result, timeout=timeout, cls=cls)
+    await download_url_to_bytesio(video_url, result, timeout=timeout, max_retries=max_retries, cls=cls)
     return VideoFromFile(result)
 
 

comfy_api_nodes/util/validation_utils.py

@@ -37,63 +37,62 @@ def validate_image_dimensions(
 
 def validate_image_aspect_ratio(
     image: torch.Tensor,
-    min_aspect_ratio: Optional[float] = None,
-    max_aspect_ratio: Optional[float] = None,
-):
-    width, height = get_image_dimensions(image)
-    aspect_ratio = width / height
-
-    if min_aspect_ratio is not None and aspect_ratio < min_aspect_ratio:
-        raise ValueError(f"Image aspect ratio must be at least {min_aspect_ratio}, got {aspect_ratio}")
-    if max_aspect_ratio is not None and aspect_ratio > max_aspect_ratio:
-        raise ValueError(f"Image aspect ratio must be at most {max_aspect_ratio}, got {aspect_ratio}")
-
-
-def validate_image_aspect_ratio_range(
-    image: torch.Tensor,
-    min_ratio: tuple[float, float],  # e.g. (1, 4)
-    max_ratio: tuple[float, float],  # e.g. (4, 1)
+    min_ratio: Optional[tuple[float, float]] = None,  # e.g. (1, 4)
+    max_ratio: Optional[tuple[float, float]] = None,  # e.g. (4, 1)
     *,
     strict: bool = True,  # True -> (min, max); False -> [min, max]
 ) -> float:
-    a1, b1 = min_ratio
-    a2, b2 = max_ratio
-    if a1 <= 0 or b1 <= 0 or a2 <= 0 or b2 <= 0:
-        raise ValueError("Ratios must be positive, like (1, 4) or (4, 1).")
-    lo, hi = (a1 / b1), (a2 / b2)
-    if lo > hi:
-        lo, hi = hi, lo
-        a1, b1, a2, b2 = a2, b2, a1, b1  # swap only for error text
+    """Validates that image aspect ratio is within min and max. If a bound is None, that side is not checked."""
     w, h = get_image_dimensions(image)
     if w <= 0 or h <= 0:
         raise ValueError(f"Invalid image dimensions: {w}x{h}")
     ar = w / h
-    ok = (lo < ar < hi) if strict else (lo <= ar <= hi)
-    if not ok:
-        op = "<" if strict else "≤"
-        raise ValueError(f"Image aspect ratio {ar:.6g} is outside allowed range: {a1}:{b1} {op} ratio {op} {a2}:{b2}")
+    _assert_ratio_bounds(ar, min_ratio=min_ratio, max_ratio=max_ratio, strict=strict)
     return ar
 
 
-def validate_aspect_ratio_closeness(
-    start_img,
-    end_img,
-    min_rel: float,
-    max_rel: float,
+def validate_images_aspect_ratio_closeness(
+    first_image: torch.Tensor,
+    second_image: torch.Tensor,
+    min_rel: float,   # e.g. 0.8
+    max_rel: float,   # e.g. 1.25
     *,
-    strict: bool = False,  # True => exclusive, False => inclusive
-) -> None:
-    w1, h1 = get_image_dimensions(start_img)
-    w2, h2 = get_image_dimensions(end_img)
+    strict: bool = False,  # True -> (min, max); False -> [min, max]
+) -> float:
+    """
+    Validates that the two images' aspect ratios are 'close'.
+    The closeness factor is C = max(ar1, ar2) / min(ar1, ar2)  (C >= 1).
+    We require C <= limit, where limit = max(max_rel, 1.0 / min_rel).
+
+    Returns the computed closeness factor C.
+    """
+    w1, h1 = get_image_dimensions(first_image)
+    w2, h2 = get_image_dimensions(second_image)
     if min(w1, h1, w2, h2) <= 0:
         raise ValueError("Invalid image dimensions")
     ar1 = w1 / h1
     ar2 = w2 / h2
-    # Normalize so it is symmetric (no need to check both ar1/ar2 and ar2/ar1)
     closeness = max(ar1, ar2) / min(ar1, ar2)
-    limit = max(max_rel, 1.0 / min_rel)  # for 0.8..1.25 this is 1.25
+    limit = max(max_rel, 1.0 / min_rel)
     if (closeness >= limit) if strict else (closeness > limit):
-        raise ValueError(f"Aspect ratios must be close: start/end={ar1/ar2:.4f}, allowed range {min_rel}–{max_rel}.")
+        raise ValueError(
+            f"Aspect ratios must be close: ar1/ar2={ar1/ar2:.2g}, "
+            f"allowed range {min_rel}–{max_rel} (limit {limit:.2g})."
+        )
+    return closeness
+
+
+def validate_aspect_ratio_string(
+    aspect_ratio: str,
+    min_ratio: Optional[tuple[float, float]] = None,  # e.g. (1, 4)
+    max_ratio: Optional[tuple[float, float]] = None,  # e.g. (4, 1)
+    *,
+    strict: bool = False,  # True -> (min, max); False -> [min, max]
+) -> float:
+    """Parses 'X:Y' and validates it against optional bounds. Returns the numeric ratio."""
+    ar = _parse_aspect_ratio_string(aspect_ratio)
+    _assert_ratio_bounds(ar, min_ratio=min_ratio, max_ratio=max_ratio, strict=strict)
+    return ar
 
 
 def validate_video_dimensions(
@@ -183,3 +182,49 @@ def validate_container_format_is_mp4(video: VideoInput) -> None:
     container_format = video.get_container_format()
     if container_format not in ["mp4", "mov,mp4,m4a,3gp,3g2,mj2"]:
         raise ValueError(f"Only MP4 container format supported. Got: {container_format}")
+
+
+def _ratio_from_tuple(r: tuple[float, float]) -> float:
+    a, b = r
+    if a <= 0 or b <= 0:
+        raise ValueError(f"Ratios must be positive, got {a}:{b}.")
+    return a / b
+
+
+def _assert_ratio_bounds(
+    ar: float,
+    *,
+    min_ratio: Optional[tuple[float, float]] = None,
+    max_ratio: Optional[tuple[float, float]] = None,
+    strict: bool = True,
+) -> None:
+    """Validate a numeric aspect ratio against optional min/max ratio bounds."""
+    lo = _ratio_from_tuple(min_ratio) if min_ratio is not None else None
+    hi = _ratio_from_tuple(max_ratio) if max_ratio is not None else None
+
+    if lo is not None and hi is not None and lo > hi:
+        lo, hi = hi, lo  # normalize order if caller swapped them
+
+    if lo is not None:
+        if (ar <= lo) if strict else (ar < lo):
+            op = "<" if strict else "≤"
+            raise ValueError(f"Aspect ratio `{ar:.2g}` must be {op} {lo:.2g}.")
+    if hi is not None:
+        if (ar >= hi) if strict else (ar > hi):
+            op = "<" if strict else "≤"
+            raise ValueError(f"Aspect ratio `{ar:.2g}` must be {op} {hi:.2g}.")
+
+
+def _parse_aspect_ratio_string(ar_str: str) -> float:
+    """Parse 'X:Y' with integer parts into a positive float ratio X/Y."""
+    parts = ar_str.split(":")
+    if len(parts) != 2:
+        raise ValueError(f"Aspect ratio must be 'X:Y' (e.g., 16:9), got '{ar_str}'.")
+    try:
+        a = int(parts[0].strip())
+        b = int(parts[1].strip())
+    except ValueError as exc:
+        raise ValueError(f"Aspect ratio must contain integers separated by ':', got '{ar_str}'.") from exc
+    if a <= 0 or b <= 0:
+        raise ValueError(f"Aspect ratio parts must be positive integers, got {a}:{b}.")
+    return a / b

comfy_execution/caching.py

@@ -1,4 +1,9 @@
+import bisect
+import gc
 import itertools
+import psutil
+import time
+import torch
 from typing import Sequence, Mapping, Dict
 from comfy_execution.graph import DynamicPrompt
 from abc import ABC, abstractmethod
@@ -48,7 +53,7 @@ class Unhashable:
 def to_hashable(obj):
     # So that we don't infinitely recurse since frozenset and tuples
     # are Sequences.
-    if isinstance(obj, (int, float, str, bool, type(None))):
+    if isinstance(obj, (int, float, str, bool, bytes, type(None))):
         return obj
     elif isinstance(obj, Mapping):
         return frozenset([(to_hashable(k), to_hashable(v)) for k, v in sorted(obj.items())])
@@ -188,6 +193,9 @@ class BasicCache:
         self._clean_cache()
         self._clean_subcaches()
 
+    def poll(self, **kwargs):
+        pass
+
     def _set_immediate(self, node_id, value):
         assert self.initialized
         cache_key = self.cache_key_set.get_data_key(node_id)
@@ -276,6 +284,9 @@ class NullCache:
     def clean_unused(self):
         pass
 
+    def poll(self, **kwargs):
+        pass
+
     def get(self, node_id):
         return None
 
@@ -336,3 +347,75 @@ class LRUCache(BasicCache):
             self._mark_used(child_id)
             self.children[cache_key].append(self.cache_key_set.get_data_key(child_id))
         return self
+
+
+#Iterating the cache for usage analysis might be expensive, so if we trigger make sure
+#to take a chunk out to give breathing space on high-node / low-ram-per-node flows.
+
+RAM_CACHE_HYSTERESIS = 1.1
+
+#This is kinda in GB but not really. It needs to be non-zero for the below heuristic
+#and as long as Multi GB models dwarf this it will approximate OOM scoring OK
+
+RAM_CACHE_DEFAULT_RAM_USAGE = 0.1
+
+#Exponential bias towards evicting older workflows so garbage will be taken out
+#in constantly changing setups.
+
+RAM_CACHE_OLD_WORKFLOW_OOM_MULTIPLIER = 1.3
+
+class RAMPressureCache(LRUCache):
+
+    def __init__(self, key_class):
+        super().__init__(key_class, 0)
+        self.timestamps = {}
+
+    def clean_unused(self):
+        self._clean_subcaches()
+
+    def set(self, node_id, value):
+        self.timestamps[self.cache_key_set.get_data_key(node_id)] = time.time()
+        super().set(node_id, value)
+
+    def get(self, node_id):
+        self.timestamps[self.cache_key_set.get_data_key(node_id)] = time.time()
+        return super().get(node_id)
+
+    def poll(self, ram_headroom):
+        def _ram_gb():
+            return psutil.virtual_memory().available / (1024**3)
+
+        if _ram_gb() > ram_headroom:
+            return
+        gc.collect()
+        if _ram_gb() > ram_headroom:
+            return
+
+        clean_list = []
+
+        for key, (outputs, _), in self.cache.items():
+            oom_score =  RAM_CACHE_OLD_WORKFLOW_OOM_MULTIPLIER ** (self.generation - self.used_generation[key])
+
+            ram_usage = RAM_CACHE_DEFAULT_RAM_USAGE
+            def scan_list_for_ram_usage(outputs):
+                nonlocal ram_usage
+                for output in outputs:
+                    if isinstance(output, list):
+                        scan_list_for_ram_usage(output)
+                    elif isinstance(output, torch.Tensor) and output.device.type == 'cpu':
+                        #score Tensors at a 50% discount for RAM usage as they are likely to
+                        #be high value intermediates
+                        ram_usage += (output.numel() * output.element_size()) * 0.5
+                    elif hasattr(output, "get_ram_usage"):
+                        ram_usage += output.get_ram_usage()
+            scan_list_for_ram_usage(outputs)
+
+            oom_score *= ram_usage
+            #In the case where we have no information on the node ram usage at all,
+            #break OOM score ties on the last touch timestamp (pure LRU)
+            bisect.insort(clean_list, (oom_score, self.timestamps[key], key))
+
+        while _ram_gb() < ram_headroom * RAM_CACHE_HYSTERESIS and clean_list:
+            _, _, key = clean_list.pop()
+            del self.cache[key]
+            gc.collect()

comfy_execution/graph.py

@@ -209,10 +209,15 @@ class ExecutionList(TopologicalSort):
             self.execution_cache_listeners[from_node_id] = set()
         self.execution_cache_listeners[from_node_id].add(to_node_id)
 
-    def get_output_cache(self, from_node_id, to_node_id):
+    def get_cache(self, from_node_id, to_node_id):
         if not to_node_id in self.execution_cache:
             return None
-        return self.execution_cache[to_node_id].get(from_node_id)
+        value = self.execution_cache[to_node_id].get(from_node_id)
+        if value is None:
+            return None
+        #Write back to the main cache on touch.
+        self.output_cache.set(from_node_id, value)
+        return value
 
     def cache_update(self, node_id, value):
         if node_id in self.execution_cache_listeners:

comfy_execution/validation.py

@@ -23,6 +23,13 @@ def validate_node_input(
     if not received_type != input_type:
         return True
 
+    # If input_type is a Combo, frontend permits a Combo output to be connected,
+    # but it is defined as a tuple of values with V3 schema.
+    # This probably should be dealt with sending one thing to the frontend and another to the backend,
+    # but this will do for now.
+    if input_type == "COMBO" and isinstance(received_type, tuple):
+        return True
+
     # Not equal, and not strings
     if not isinstance(received_type, str) or not isinstance(input_type, str):
         return False

comfy_extras/nodes_flipflop.py

@@ -1,43 +0,0 @@
-from __future__ import annotations
-from typing_extensions import override
-
-from comfy_api.latest import ComfyExtension, io
-
-
-class FlipFlop(io.ComfyNode):
-    @classmethod
-    def define_schema(cls) -> io.Schema:
-        return io.Schema(
-            node_id="FlipFlopNew",
-            display_name="FlipFlop (New)",
-            category="_for_testing",
-            inputs=[
-                io.Model.Input(id="model"),
-                io.Float.Input(id="block_percentage", default=1.0, min=0.0, max=1.0, step=0.01),
-            ],
-            outputs=[
-                io.Model.Output()
-            ],
-            description="Apply FlipFlop transformation to model using setup_flipflop_holders method"
-        )
-
-    @classmethod
-    def execute(cls, model: io.Model.Type, block_percentage: float) -> io.NodeOutput:
-        # NOTE: this is just a hacky prototype still, this would not be exposed as a node.
-        # At the moment, this modifies the underlying model with no way to 'unpatch' it.
-        model = model.clone()
-        if not hasattr(model.model.diffusion_model, "setup_flipflop_holders"):
-            raise ValueError("Model does not have flipflop holders; FlipFlop not supported")
-        model.model.diffusion_model.setup_flipflop_holders(block_percentage)
-        return io.NodeOutput(model)
-
-class FlipFlopExtension(ComfyExtension):
-    @override
-    async def get_node_list(self) -> list[type[io.ComfyNode]]:
-        return [
-            FlipFlop,
-        ]
-
-
-async def comfy_entrypoint() -> FlipFlopExtension:
-    return FlipFlopExtension()

comfy_extras/nodes_hypernetwork.py

@@ -2,6 +2,9 @@ import comfy.utils
 import folder_paths
 import torch
 import logging
+from comfy_api.latest import IO, ComfyExtension
+from typing_extensions import override
+
 
 def load_hypernetwork_patch(path, strength):
     sd = comfy.utils.load_torch_file(path, safe_load=True)
@@ -94,27 +97,42 @@ def load_hypernetwork_patch(path, strength):
 
     return hypernetwork_patch(out, strength)
 
-class HypernetworkLoader:
+class HypernetworkLoader(IO.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": { "model": ("MODEL",),
-                              "hypernetwork_name": (folder_paths.get_filename_list("hypernetworks"), ),
-                              "strength": ("FLOAT", {"default": 1.0, "min": -10.0, "max": 10.0, "step": 0.01}),
-                              }}
-    RETURN_TYPES = ("MODEL",)
-    FUNCTION = "load_hypernetwork"
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="HypernetworkLoader",
+            category="loaders",
+            inputs=[
+                IO.Model.Input("model"),
+                IO.Combo.Input("hypernetwork_name", options=folder_paths.get_filename_list("hypernetworks")),
+                IO.Float.Input("strength", default=1.0, min=-10.0, max=10.0, step=0.01),
+            ],
+            outputs=[
+                IO.Model.Output(),
+            ],
+        )
 
-    CATEGORY = "loaders"
-
-    def load_hypernetwork(self, model, hypernetwork_name, strength):
+    @classmethod
+    def execute(cls, model, hypernetwork_name, strength) -> IO.NodeOutput:
         hypernetwork_path = folder_paths.get_full_path_or_raise("hypernetworks", hypernetwork_name)
         model_hypernetwork = model.clone()
         patch = load_hypernetwork_patch(hypernetwork_path, strength)
         if patch is not None:
             model_hypernetwork.set_model_attn1_patch(patch)
             model_hypernetwork.set_model_attn2_patch(patch)
-        return (model_hypernetwork,)
+        return IO.NodeOutput(model_hypernetwork)
 
-NODE_CLASS_MAPPINGS = {
-    "HypernetworkLoader": HypernetworkLoader
-}
+    load_hypernetwork = execute  # TODO: remove
+
+
+class HyperNetworkExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[IO.ComfyNode]]:
+        return [
+            HypernetworkLoader,
+        ]
+
+
+async def comfy_entrypoint() -> HyperNetworkExtension:
+    return HyperNetworkExtension()

comfy_extras/nodes_rope.py

@@ -0,0 +1,47 @@
+from comfy_api.latest import ComfyExtension, io
+from typing_extensions import override
+
+
+class ScaleROPE(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="ScaleROPE",
+            category="advanced/model_patches",
+            description="Scale and shift the ROPE of the model.",
+            is_experimental=True,
+            inputs=[
+                io.Model.Input("model"),
+                io.Float.Input("scale_x", default=1.0, min=0.0, max=100.0, step=0.1),
+                io.Float.Input("shift_x", default=0.0, min=-256.0, max=256.0, step=0.1),
+
+                io.Float.Input("scale_y", default=1.0, min=0.0, max=100.0, step=0.1),
+                io.Float.Input("shift_y", default=0.0, min=-256.0, max=256.0, step=0.1),
+
+                io.Float.Input("scale_t", default=1.0, min=0.0, max=100.0, step=0.1),
+                io.Float.Input("shift_t", default=0.0, min=-256.0, max=256.0, step=0.1),
+
+
+            ],
+            outputs=[
+                io.Model.Output(),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, model, scale_x, shift_x, scale_y, shift_y, scale_t, shift_t) -> io.NodeOutput:
+        m = model.clone()
+        m.set_model_rope_options(scale_x, shift_x, scale_y, shift_y, scale_t, shift_t)
+        return io.NodeOutput(m)
+
+
+class RopeExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            ScaleROPE
+        ]
+
+
+async def comfy_entrypoint() -> RopeExtension:
+    return RopeExtension()

execution.py

@@ -21,6 +21,7 @@ from comfy_execution.caching import (
     NullCache,
     HierarchicalCache,
     LRUCache,
+    RAMPressureCache,
 )
 from comfy_execution.graph import (
     DynamicPrompt,
@@ -88,49 +89,56 @@ class IsChangedCache:
         return self.is_changed[node_id]
 
 
+class CacheEntry(NamedTuple):
+    ui: dict
+    outputs: list
+
+
 class CacheType(Enum):
     CLASSIC = 0
     LRU = 1
     NONE = 2
+    RAM_PRESSURE = 3
 
 
 class CacheSet:
-    def __init__(self, cache_type=None, cache_size=None):
+    def __init__(self, cache_type=None, cache_args={}):
         if cache_type == CacheType.NONE:
             self.init_null_cache()
             logging.info("Disabling intermediate node cache.")
+        elif cache_type == CacheType.RAM_PRESSURE:
+            cache_ram = cache_args.get("ram", 16.0)
+            self.init_ram_cache(cache_ram)
+            logging.info("Using RAM pressure cache.")
         elif cache_type == CacheType.LRU:
-            if cache_size is None:
-                cache_size = 0
+            cache_size = cache_args.get("lru", 0)
             self.init_lru_cache(cache_size)
             logging.info("Using LRU cache")
         else:
             self.init_classic_cache()
 
-        self.all = [self.outputs, self.ui, self.objects]
+        self.all = [self.outputs, self.objects]
 
     # Performs like the old cache -- dump data ASAP
     def init_classic_cache(self):
         self.outputs = HierarchicalCache(CacheKeySetInputSignature)
-        self.ui = HierarchicalCache(CacheKeySetInputSignature)
         self.objects = HierarchicalCache(CacheKeySetID)
 
     def init_lru_cache(self, cache_size):
         self.outputs = LRUCache(CacheKeySetInputSignature, max_size=cache_size)
-        self.ui = LRUCache(CacheKeySetInputSignature, max_size=cache_size)
+        self.objects = HierarchicalCache(CacheKeySetID)
+
+    def init_ram_cache(self, min_headroom):
+        self.outputs = RAMPressureCache(CacheKeySetInputSignature)
         self.objects = HierarchicalCache(CacheKeySetID)
 
     def init_null_cache(self):
         self.outputs = NullCache()
-        #The UI cache is expected to be iterable at the end of each workflow
-        #so it must cache at least a full workflow. Use Heirachical
-        self.ui = HierarchicalCache(CacheKeySetInputSignature)
         self.objects = NullCache()
 
     def recursive_debug_dump(self):
         result = {
             "outputs": self.outputs.recursive_debug_dump(),
-            "ui": self.ui.recursive_debug_dump(),
         }
         return result
 
@@ -157,14 +165,14 @@ def get_input_data(inputs, class_def, unique_id, execution_list=None, dynprompt=
             if execution_list is None:
                 mark_missing()
                 continue # This might be a lazily-evaluated input
-            cached_output = execution_list.get_output_cache(input_unique_id, unique_id)
-            if cached_output is None:
+            cached = execution_list.get_cache(input_unique_id, unique_id)
+            if cached is None or cached.outputs is None:
                 mark_missing()
                 continue
-            if output_index >= len(cached_output):
+            if output_index >= len(cached.outputs):
                 mark_missing()
                 continue
-            obj = cached_output[output_index]
+            obj = cached.outputs[output_index]
             input_data_all[x] = obj
         elif input_category is not None:
             input_data_all[x] = [input_data]
@@ -393,7 +401,7 @@ def format_value(x):
     else:
         return str(x)
 
-async def execute(server, dynprompt, caches, current_item, extra_data, executed, prompt_id, execution_list, pending_subgraph_results, pending_async_nodes):
+async def execute(server, dynprompt, caches, current_item, extra_data, executed, prompt_id, execution_list, pending_subgraph_results, pending_async_nodes, ui_outputs):
     unique_id = current_item
     real_node_id = dynprompt.get_real_node_id(unique_id)
     display_node_id = dynprompt.get_display_node_id(unique_id)
@@ -401,12 +409,15 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed,
     inputs = dynprompt.get_node(unique_id)['inputs']
     class_type = dynprompt.get_node(unique_id)['class_type']
     class_def = nodes.NODE_CLASS_MAPPINGS[class_type]
-    if caches.outputs.get(unique_id) is not None:
+    cached = caches.outputs.get(unique_id)
+    if cached is not None:
         if server.client_id is not None:
-            cached_output = caches.ui.get(unique_id) or {}
-            server.send_sync("executed", { "node": unique_id, "display_node": display_node_id, "output": cached_output.get("output",None), "prompt_id": prompt_id }, server.client_id)
+            cached_ui = cached.ui or {}
+            server.send_sync("executed", { "node": unique_id, "display_node": display_node_id, "output": cached_ui.get("output",None), "prompt_id": prompt_id }, server.client_id)
+            if cached.ui is not None:
+                ui_outputs[unique_id] = cached.ui
         get_progress_state().finish_progress(unique_id)
-        execution_list.cache_update(unique_id, caches.outputs.get(unique_id))
+        execution_list.cache_update(unique_id, cached)
         return (ExecutionResult.SUCCESS, None, None)
 
     input_data_all = None
@@ -436,8 +447,8 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed,
                     for r in result:
                         if is_link(r):
                             source_node, source_output = r[0], r[1]
-                            node_output = execution_list.get_output_cache(source_node, unique_id)[source_output]
-                            for o in node_output:
+                            node_cached = execution_list.get_cache(source_node, unique_id)
+                            for o in node_cached.outputs[source_output]:
                                 resolved_output.append(o)
 
                         else:
@@ -507,7 +518,7 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed,
                 asyncio.create_task(await_completion())
                 return (ExecutionResult.PENDING, None, None)
         if len(output_ui) > 0:
-            caches.ui.set(unique_id, {
+            ui_outputs[unique_id] = {
                 "meta": {
                     "node_id": unique_id,
                     "display_node": display_node_id,
@@ -515,7 +526,7 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed,
                     "real_node_id": real_node_id,
                 },
                 "output": output_ui
-            })
+            }
             if server.client_id is not None:
                 server.send_sync("executed", { "node": unique_id, "display_node": display_node_id, "output": output_ui, "prompt_id": prompt_id }, server.client_id)
         if has_subgraph:
@@ -554,8 +565,9 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed,
             pending_subgraph_results[unique_id] = cached_outputs
             return (ExecutionResult.PENDING, None, None)
 
-        caches.outputs.set(unique_id, output_data)
-        execution_list.cache_update(unique_id, output_data)
+        cache_entry = CacheEntry(ui=ui_outputs.get(unique_id), outputs=output_data)
+        execution_list.cache_update(unique_id, cache_entry)
+        caches.outputs.set(unique_id, cache_entry)
 
     except comfy.model_management.InterruptProcessingException as iex:
         logging.info("Processing interrupted")
@@ -600,14 +612,14 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed,
     return (ExecutionResult.SUCCESS, None, None)
 
 class PromptExecutor:
-    def __init__(self, server, cache_type=False, cache_size=None):
-        self.cache_size = cache_size
+    def __init__(self, server, cache_type=False, cache_args=None):
+        self.cache_args = cache_args
         self.cache_type = cache_type
         self.server = server
         self.reset()
 
     def reset(self):
-        self.caches = CacheSet(cache_type=self.cache_type, cache_size=self.cache_size)
+        self.caches = CacheSet(cache_type=self.cache_type, cache_args=self.cache_args)
         self.status_messages = []
         self.success = True
 
@@ -682,6 +694,7 @@ class PromptExecutor:
                           broadcast=False)
             pending_subgraph_results = {}
             pending_async_nodes = {} # TODO - Unify this with pending_subgraph_results
+            ui_node_outputs = {}
             executed = set()
             execution_list = ExecutionList(dynamic_prompt, self.caches.outputs)
             current_outputs = self.caches.outputs.all_node_ids()
@@ -695,7 +708,7 @@ class PromptExecutor:
                     break
 
                 assert node_id is not None, "Node ID should not be None at this point"
-                result, error, ex = await execute(self.server, dynamic_prompt, self.caches, node_id, extra_data, executed, prompt_id, execution_list, pending_subgraph_results, pending_async_nodes)
+                result, error, ex = await execute(self.server, dynamic_prompt, self.caches, node_id, extra_data, executed, prompt_id, execution_list, pending_subgraph_results, pending_async_nodes, ui_node_outputs)
                 self.success = result != ExecutionResult.FAILURE
                 if result == ExecutionResult.FAILURE:
                     self.handle_execution_error(prompt_id, dynamic_prompt.original_prompt, current_outputs, executed, error, ex)
@@ -704,18 +717,16 @@ class PromptExecutor:
                     execution_list.unstage_node_execution()
                 else: # result == ExecutionResult.SUCCESS:
                     execution_list.complete_node_execution()
+                self.caches.outputs.poll(ram_headroom=self.cache_args["ram"])
             else:
                 # Only execute when the while-loop ends without break
                 self.add_message("execution_success", { "prompt_id": prompt_id }, broadcast=False)
 
             ui_outputs = {}
             meta_outputs = {}
-            all_node_ids = self.caches.ui.all_node_ids()
-            for node_id in all_node_ids:
-                ui_info = self.caches.ui.get(node_id)
-                if ui_info is not None:
-                    ui_outputs[node_id] = ui_info["output"]
-                    meta_outputs[node_id] = ui_info["meta"]
+            for node_id, ui_info in ui_node_outputs.items():
+                ui_outputs[node_id] = ui_info["output"]
+                meta_outputs[node_id] = ui_info["meta"]
             self.history_result = {
                 "outputs": ui_outputs,
                 "meta": meta_outputs,

main.py

@@ -172,10 +172,12 @@ def prompt_worker(q, server_instance):
     cache_type = execution.CacheType.CLASSIC
     if args.cache_lru > 0:
         cache_type = execution.CacheType.LRU
+    elif args.cache_ram > 0:
+        cache_type = execution.CacheType.RAM_PRESSURE
     elif args.cache_none:
         cache_type = execution.CacheType.NONE
 
-    e = execution.PromptExecutor(server_instance, cache_type=cache_type, cache_size=args.cache_lru)
+    e = execution.PromptExecutor(server_instance, cache_type=cache_type, cache_args={ "lru" : args.cache_lru, "ram" : args.cache_ram } )
     last_gc_collect = 0
     need_gc = False
     gc_collect_interval = 10.0

nodes.py

@@ -2329,7 +2329,7 @@ async def init_builtin_extra_nodes():
         "nodes_model_patch.py",
         "nodes_easycache.py",
         "nodes_audio_encoder.py",
-        "nodes_flipflop.py",
+        "nodes_rope.py",
     ]
 
     import_failed = []

requirements.txt

@@ -1,6 +1,6 @@
 comfyui-frontend-package==1.28.8
-comfyui-workflow-templates==0.2.4
-comfyui-embedded-docs==0.3.0
+comfyui-workflow-templates==0.2.11
+comfyui-embedded-docs==0.3.1
 torch
 torchsde
 torchvision

tests-unit/comfy_quant/test_mixed_precision.py

@@ -14,7 +14,7 @@ if not has_gpu():
     args.cpu = True
 
 from comfy import ops
-from comfy.quant_ops import QuantizedTensor, TensorCoreFP8Layout
+from comfy.quant_ops import QuantizedTensor
 
 
 class SimpleModel(torch.nn.Module):
@@ -104,14 +104,14 @@ class TestMixedPrecisionOps(unittest.TestCase):
 
         # Verify weights are wrapped in QuantizedTensor
         self.assertIsInstance(model.layer1.weight, QuantizedTensor)
-        self.assertEqual(model.layer1.weight._layout_type, TensorCoreFP8Layout)
+        self.assertEqual(model.layer1.weight._layout_type, "TensorCoreFP8Layout")
 
         # Layer 2 should NOT be quantized
         self.assertNotIsInstance(model.layer2.weight, QuantizedTensor)
 
         # Layer 3 should be quantized
         self.assertIsInstance(model.layer3.weight, QuantizedTensor)
-        self.assertEqual(model.layer3.weight._layout_type, TensorCoreFP8Layout)
+        self.assertEqual(model.layer3.weight._layout_type, "TensorCoreFP8Layout")
 
         # Verify scales were loaded
         self.assertEqual(model.layer1.weight._layout_params['scale'].item(), 2.0)
@@ -155,7 +155,7 @@ class TestMixedPrecisionOps(unittest.TestCase):
         # Verify layer1.weight is a QuantizedTensor with scale preserved
         self.assertIsInstance(state_dict2["layer1.weight"], QuantizedTensor)
         self.assertEqual(state_dict2["layer1.weight"]._layout_params['scale'].item(), 3.0)
-        self.assertEqual(state_dict2["layer1.weight"]._layout_type, TensorCoreFP8Layout)
+        self.assertEqual(state_dict2["layer1.weight"]._layout_type, "TensorCoreFP8Layout")
 
         # Verify non-quantized layers are standard tensors
         self.assertNotIsInstance(state_dict2["layer2.weight"], QuantizedTensor)

tests-unit/comfy_quant/test_quant_registry.py

@@ -25,14 +25,14 @@ class TestQuantizedTensor(unittest.TestCase):
         scale = torch.tensor(2.0)
         layout_params = {'scale': scale, 'orig_dtype': torch.bfloat16}
 
-        qt = QuantizedTensor(fp8_data, TensorCoreFP8Layout, layout_params)
+        qt = QuantizedTensor(fp8_data, "TensorCoreFP8Layout", layout_params)
 
         self.assertIsInstance(qt, QuantizedTensor)
         self.assertEqual(qt.shape, (256, 128))
         self.assertEqual(qt.dtype, torch.float8_e4m3fn)
         self.assertEqual(qt._layout_params['scale'], scale)
         self.assertEqual(qt._layout_params['orig_dtype'], torch.bfloat16)
-        self.assertEqual(qt._layout_type, TensorCoreFP8Layout)
+        self.assertEqual(qt._layout_type, "TensorCoreFP8Layout")
 
     def test_dequantize(self):
         """Test explicit dequantization"""
@@ -41,7 +41,7 @@ class TestQuantizedTensor(unittest.TestCase):
         scale = torch.tensor(3.0)
         layout_params = {'scale': scale, 'orig_dtype': torch.float32}
 
-        qt = QuantizedTensor(fp8_data, TensorCoreFP8Layout, layout_params)
+        qt = QuantizedTensor(fp8_data, "TensorCoreFP8Layout", layout_params)
         dequantized = qt.dequantize()
 
         self.assertEqual(dequantized.dtype, torch.float32)
@@ -54,7 +54,7 @@ class TestQuantizedTensor(unittest.TestCase):
 
         qt = QuantizedTensor.from_float(
             float_tensor,
-            TensorCoreFP8Layout,
+            "TensorCoreFP8Layout",
             scale=scale,
             dtype=torch.float8_e4m3fn
         )
@@ -77,28 +77,28 @@ class TestGenericUtilities(unittest.TestCase):
         fp8_data = torch.randn(10, 20, dtype=torch.float32).to(torch.float8_e4m3fn)
         scale = torch.tensor(1.5)
         layout_params = {'scale': scale, 'orig_dtype': torch.float32}
-        qt = QuantizedTensor(fp8_data, TensorCoreFP8Layout, layout_params)
+        qt = QuantizedTensor(fp8_data, "TensorCoreFP8Layout", layout_params)
 
         # Detach should return a new QuantizedTensor
         qt_detached = qt.detach()
 
         self.assertIsInstance(qt_detached, QuantizedTensor)
         self.assertEqual(qt_detached.shape, qt.shape)
-        self.assertEqual(qt_detached._layout_type, TensorCoreFP8Layout)
+        self.assertEqual(qt_detached._layout_type, "TensorCoreFP8Layout")
 
     def test_clone(self):
         """Test clone operation on quantized tensor"""
         fp8_data = torch.randn(10, 20, dtype=torch.float32).to(torch.float8_e4m3fn)
         scale = torch.tensor(1.5)
         layout_params = {'scale': scale, 'orig_dtype': torch.float32}
-        qt = QuantizedTensor(fp8_data, TensorCoreFP8Layout, layout_params)
+        qt = QuantizedTensor(fp8_data, "TensorCoreFP8Layout", layout_params)
 
         # Clone should return a new QuantizedTensor
         qt_cloned = qt.clone()
 
         self.assertIsInstance(qt_cloned, QuantizedTensor)
         self.assertEqual(qt_cloned.shape, qt.shape)
-        self.assertEqual(qt_cloned._layout_type, TensorCoreFP8Layout)
+        self.assertEqual(qt_cloned._layout_type, "TensorCoreFP8Layout")
 
         # Verify it's a deep copy
         self.assertIsNot(qt_cloned._qdata, qt._qdata)
@@ -109,7 +109,7 @@ class TestGenericUtilities(unittest.TestCase):
         fp8_data = torch.randn(10, 20, dtype=torch.float32).to(torch.float8_e4m3fn)
         scale = torch.tensor(1.5)
         layout_params = {'scale': scale, 'orig_dtype': torch.float32}
-        qt = QuantizedTensor(fp8_data, TensorCoreFP8Layout, layout_params)
+        qt = QuantizedTensor(fp8_data, "TensorCoreFP8Layout", layout_params)
 
         # Moving to same device should work (CPU to CPU)
         qt_cpu = qt.to('cpu')
@@ -169,7 +169,7 @@ class TestFallbackMechanism(unittest.TestCase):
         scale = torch.tensor(1.0)
         a_q = QuantizedTensor.from_float(
             a_fp32,
-            TensorCoreFP8Layout,
+            "TensorCoreFP8Layout",
             scale=scale,
             dtype=torch.float8_e4m3fn
         )