ComfyUI v0.3.74

It looks like the synchronous version of the public API broke due to an
addition of `from __future__ import annotations`. This change updates
the async-to-sync adapter to work with both types of type annotations.

README.md

@@ -67,6 +67,7 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
    - [HiDream](https://comfyanonymous.github.io/ComfyUI_examples/hidream/)
    - [Qwen Image](https://comfyanonymous.github.io/ComfyUI_examples/qwen_image/)
    - [Hunyuan Image 2.1](https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_image/)
+   - [Flux 2](https://comfyanonymous.github.io/ComfyUI_examples/flux2/)
 - Image Editing Models
    - [Omnigen 2](https://comfyanonymous.github.io/ComfyUI_examples/omnigen/)
    - [Flux Kontext](https://comfyanonymous.github.io/ComfyUI_examples/flux/#flux-kontext-image-editing-model)

comfy/cldm/cldm.py

@@ -413,7 +413,8 @@ class ControlNet(nn.Module):
         out_middle = []
 
         if self.num_classes is not None:
-            assert y.shape[0] == x.shape[0]
+            if y is None:
+                raise ValueError("y is None, did you try using a controlnet for SDXL on SD1?")
             emb = emb + self.label_emb(y)
 
         h = x

comfy/cli_args.py

@@ -160,7 +160,7 @@ parser.add_argument("--windows-standalone-build", action="store_true", help="Win
 parser.add_argument("--disable-metadata", action="store_true", help="Disable saving prompt metadata in files.")
 parser.add_argument("--disable-all-custom-nodes", action="store_true", help="Disable loading all custom nodes.")
 parser.add_argument("--whitelist-custom-nodes", type=str, nargs='+', default=[], help="Specify custom node folders to load even when --disable-all-custom-nodes is enabled.")
-parser.add_argument("--disable-api-nodes", action="store_true", help="Disable loading all api nodes.")
+parser.add_argument("--disable-api-nodes", action="store_true", help="Disable loading all api nodes. Also prevents the frontend from communicating with the internet.")
 
 parser.add_argument("--multi-user", action="store_true", help="Enables per-user storage.")
 

comfy/latent_formats.py

@@ -178,6 +178,15 @@ class Flux(SD3):
     def process_out(self, latent):
         return (latent / self.scale_factor) + self.shift_factor
 
+class Flux2(LatentFormat):
+    latent_channels = 128
+
+    def process_in(self, latent):
+        return latent
+
+    def process_out(self, latent):
+        return latent
+
 class Mochi(LatentFormat):
     latent_channels = 12
     latent_dimensions = 3
@@ -611,6 +620,66 @@ class HunyuanImage21Refiner(LatentFormat):
     latent_dimensions = 3
     scale_factor = 1.03682
 
+    def process_in(self, latent):
+        out = latent * self.scale_factor
+        out = torch.cat((out[:, :, :1], out), dim=2)
+        out = out.permute(0, 2, 1, 3, 4)
+        b, f_times_2, c, h, w = out.shape
+        out = out.reshape(b, f_times_2 // 2, 2 * c, h, w)
+        out = out.permute(0, 2, 1, 3, 4).contiguous()
+        return out
+
+    def process_out(self, latent):
+        z = latent / self.scale_factor
+        z = z.permute(0, 2, 1, 3, 4)
+        b, f, c, h, w = z.shape
+        z = z.reshape(b, f, 2, c // 2, h, w)
+        z = z.permute(0, 1, 2, 3, 4, 5).reshape(b, f * 2, c // 2, h, w)
+        z = z.permute(0, 2, 1, 3, 4)
+        z = z[:, :, 1:]
+        return z
+
+class HunyuanVideo15(LatentFormat):
+    latent_rgb_factors = [
+        [ 0.0568, -0.0521, -0.0131],
+        [ 0.0014,  0.0735,  0.0326],
+        [ 0.0186,  0.0531, -0.0138],
+        [-0.0031,  0.0051,  0.0288],
+        [ 0.0110,  0.0556,  0.0432],
+        [-0.0041, -0.0023, -0.0485],
+        [ 0.0530,  0.0413,  0.0253],
+        [ 0.0283,  0.0251,  0.0339],
+        [ 0.0277, -0.0372, -0.0093],
+        [ 0.0393,  0.0944,  0.1131],
+        [ 0.0020,  0.0251,  0.0037],
+        [-0.0017,  0.0012,  0.0234],
+        [ 0.0468,  0.0436,  0.0203],
+        [ 0.0354,  0.0439, -0.0233],
+        [ 0.0090,  0.0123,  0.0346],
+        [ 0.0382,  0.0029,  0.0217],
+        [ 0.0261, -0.0300,  0.0030],
+        [-0.0088, -0.0220, -0.0283],
+        [-0.0272, -0.0121, -0.0363],
+        [-0.0664, -0.0622,  0.0144],
+        [ 0.0414,  0.0479,  0.0529],
+        [ 0.0355,  0.0612, -0.0247],
+        [ 0.0147,  0.0264,  0.0174],
+        [ 0.0438,  0.0038,  0.0542],
+        [ 0.0431, -0.0573, -0.0033],
+        [-0.0162, -0.0211, -0.0406],
+        [-0.0487, -0.0295, -0.0393],
+        [ 0.0005, -0.0109,  0.0253],
+        [ 0.0296,  0.0591,  0.0353],
+        [ 0.0119,  0.0181, -0.0306],
+        [-0.0085, -0.0362,  0.0229],
+        [ 0.0005, -0.0106,  0.0242]
+    ]
+
+    latent_rgb_factors_bias = [ 0.0456, -0.0202, -0.0644]
+    latent_channels = 32
+    latent_dimensions = 3
+    scale_factor = 1.03682
+
 class Hunyuan3Dv2(LatentFormat):
     latent_channels = 64
     latent_dimensions = 1

comfy/ldm/chroma/model.py

@@ -179,7 +179,10 @@ class Chroma(nn.Module):
         pe = self.pe_embedder(ids)
 
         blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.double_blocks)
+        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.double_blocks):
+            transformer_options["block_index"] = i
             if i not in self.skip_mmdit:
                 double_mod = (
                     self.get_modulations(mod_vectors, "double_img", idx=i),
@@ -222,7 +225,10 @@ class Chroma(nn.Module):
 
         img = torch.cat((txt, img), 1)
 
+        transformer_options["total_blocks"] = len(self.single_blocks)
+        transformer_options["block_type"] = "single"
         for i, block in enumerate(self.single_blocks):
+            transformer_options["block_index"] = i
             if i not in self.skip_dit:
                 single_mod = self.get_modulations(mod_vectors, "single", idx=i)
                 if ("single_block", i) in blocks_replace:

comfy/ldm/flux/layers.py

@@ -48,11 +48,11 @@ def timestep_embedding(t: Tensor, dim, max_period=10000, time_factor: float = 10
     return embedding
 
 class MLPEmbedder(nn.Module):
-    def __init__(self, in_dim: int, hidden_dim: int, dtype=None, device=None, operations=None):
+    def __init__(self, in_dim: int, hidden_dim: int, bias=True, dtype=None, device=None, operations=None):
         super().__init__()
-        self.in_layer = operations.Linear(in_dim, hidden_dim, bias=True, dtype=dtype, device=device)
+        self.in_layer = operations.Linear(in_dim, hidden_dim, bias=bias, dtype=dtype, device=device)
         self.silu = nn.SiLU()
-        self.out_layer = operations.Linear(hidden_dim, hidden_dim, bias=True, dtype=dtype, device=device)
+        self.out_layer = operations.Linear(hidden_dim, hidden_dim, bias=bias, dtype=dtype, device=device)
 
     def forward(self, x: Tensor) -> Tensor:
         return self.out_layer(self.silu(self.in_layer(x)))
@@ -80,14 +80,14 @@ class QKNorm(torch.nn.Module):
 
 
 class SelfAttention(nn.Module):
-    def __init__(self, dim: int, num_heads: int = 8, qkv_bias: bool = False, dtype=None, device=None, operations=None):
+    def __init__(self, dim: int, num_heads: int = 8, qkv_bias: bool = False, proj_bias: bool = True, dtype=None, device=None, operations=None):
         super().__init__()
         self.num_heads = num_heads
         head_dim = dim // num_heads
 
         self.qkv = operations.Linear(dim, dim * 3, bias=qkv_bias, dtype=dtype, device=device)
         self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations)
-        self.proj = operations.Linear(dim, dim, dtype=dtype, device=device)
+        self.proj = operations.Linear(dim, dim, bias=proj_bias, dtype=dtype, device=device)
 
 
 @dataclass
@@ -98,11 +98,11 @@ class ModulationOut:
 
 
 class Modulation(nn.Module):
-    def __init__(self, dim: int, double: bool, dtype=None, device=None, operations=None):
+    def __init__(self, dim: int, double: bool, bias=True, dtype=None, device=None, operations=None):
         super().__init__()
         self.is_double = double
         self.multiplier = 6 if double else 3
-        self.lin = operations.Linear(dim, self.multiplier * dim, bias=True, dtype=dtype, device=device)
+        self.lin = operations.Linear(dim, self.multiplier * dim, bias=bias, dtype=dtype, device=device)
 
     def forward(self, vec: Tensor) -> tuple:
         if vec.ndim == 2:
@@ -129,8 +129,18 @@ def apply_mod(tensor, m_mult, m_add=None, modulation_dims=None):
         return tensor
 
 
+class SiLUActivation(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.gate_fn = nn.SiLU()
+
+    def forward(self, x: Tensor) -> Tensor:
+        x1, x2 = x.chunk(2, dim=-1)
+        return self.gate_fn(x1) * x2
+
+
 class DoubleStreamBlock(nn.Module):
-    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, modulation=True, dtype=None, device=None, operations=None):
+    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, modulation=True, mlp_silu_act=False, proj_bias=True, dtype=None, device=None, operations=None):
         super().__init__()
 
         mlp_hidden_dim = int(hidden_size * mlp_ratio)
@@ -142,27 +152,44 @@ class DoubleStreamBlock(nn.Module):
             self.img_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations)
 
         self.img_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
-        self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations)
+        self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, proj_bias=proj_bias, dtype=dtype, device=device, operations=operations)
 
         self.img_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
-        self.img_mlp = nn.Sequential(
-            operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
-            nn.GELU(approximate="tanh"),
-            operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
-        )
+
+        if mlp_silu_act:
+            self.img_mlp = nn.Sequential(
+                operations.Linear(hidden_size, mlp_hidden_dim * 2, bias=False, dtype=dtype, device=device),
+                SiLUActivation(),
+                operations.Linear(mlp_hidden_dim, hidden_size, bias=False, dtype=dtype, device=device),
+            )
+        else:
+            self.img_mlp = nn.Sequential(
+                operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
+                nn.GELU(approximate="tanh"),
+                operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
+            )
 
         if self.modulation:
             self.txt_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations)
 
         self.txt_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
-        self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations)
+        self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, proj_bias=proj_bias, dtype=dtype, device=device, operations=operations)
 
         self.txt_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
-        self.txt_mlp = nn.Sequential(
-            operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
-            nn.GELU(approximate="tanh"),
-            operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
-        )
+
+        if mlp_silu_act:
+            self.txt_mlp = nn.Sequential(
+                operations.Linear(hidden_size, mlp_hidden_dim * 2, bias=False, dtype=dtype, device=device),
+                SiLUActivation(),
+                operations.Linear(mlp_hidden_dim, hidden_size, bias=False, dtype=dtype, device=device),
+            )
+        else:
+            self.txt_mlp = nn.Sequential(
+                operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
+                nn.GELU(approximate="tanh"),
+                operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
+            )
+
         self.flipped_img_txt = flipped_img_txt
 
     def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims_img=None, modulation_dims_txt=None, transformer_options={}):
@@ -246,6 +273,8 @@ class SingleStreamBlock(nn.Module):
         mlp_ratio: float = 4.0,
         qk_scale: float = None,
         modulation=True,
+        mlp_silu_act=False,
+        bias=True,
         dtype=None,
         device=None,
         operations=None
@@ -257,17 +286,24 @@ class SingleStreamBlock(nn.Module):
         self.scale = qk_scale or head_dim**-0.5
 
         self.mlp_hidden_dim = int(hidden_size * mlp_ratio)
+
+        self.mlp_hidden_dim_first = self.mlp_hidden_dim
+        if mlp_silu_act:
+            self.mlp_hidden_dim_first = int(hidden_size * mlp_ratio * 2)
+            self.mlp_act = SiLUActivation()
+        else:
+            self.mlp_act = nn.GELU(approximate="tanh")
+
         # qkv and mlp_in
-        self.linear1 = operations.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim, dtype=dtype, device=device)
+        self.linear1 = operations.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim_first, bias=bias, dtype=dtype, device=device)
         # proj and mlp_out
-        self.linear2 = operations.Linear(hidden_size + self.mlp_hidden_dim, hidden_size, dtype=dtype, device=device)
+        self.linear2 = operations.Linear(hidden_size + self.mlp_hidden_dim, hidden_size, bias=bias, dtype=dtype, device=device)
 
         self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations)
 
         self.hidden_size = hidden_size
         self.pre_norm = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
 
-        self.mlp_act = nn.GELU(approximate="tanh")
         if modulation:
             self.modulation = Modulation(hidden_size, double=False, dtype=dtype, device=device, operations=operations)
         else:
@@ -279,7 +315,7 @@ class SingleStreamBlock(nn.Module):
         else:
             mod = vec
 
-        qkv, mlp = torch.split(self.linear1(apply_mod(self.pre_norm(x), (1 + mod.scale), mod.shift, modulation_dims)), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1)
+        qkv, mlp = torch.split(self.linear1(apply_mod(self.pre_norm(x), (1 + mod.scale), mod.shift, modulation_dims)), [3 * self.hidden_size, self.mlp_hidden_dim_first], dim=-1)
 
         q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
         del qkv
@@ -298,11 +334,11 @@ class SingleStreamBlock(nn.Module):
 
 
 class LastLayer(nn.Module):
-    def __init__(self, hidden_size: int, patch_size: int, out_channels: int, dtype=None, device=None, operations=None):
+    def __init__(self, hidden_size: int, patch_size: int, out_channels: int, bias=True, dtype=None, device=None, operations=None):
         super().__init__()
         self.norm_final = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
-        self.linear = operations.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True, dtype=dtype, device=device)
-        self.adaLN_modulation = nn.Sequential(nn.SiLU(), operations.Linear(hidden_size, 2 * hidden_size, bias=True, dtype=dtype, device=device))
+        self.linear = operations.Linear(hidden_size, patch_size * patch_size * out_channels, bias=bias, dtype=dtype, device=device)
+        self.adaLN_modulation = nn.Sequential(nn.SiLU(), operations.Linear(hidden_size, 2 * hidden_size, bias=bias, dtype=dtype, device=device))
 
     def forward(self, x: Tensor, vec: Tensor, modulation_dims=None) -> Tensor:
         if vec.ndim == 2:

comfy/ldm/flux/model.py

@@ -15,6 +15,7 @@ from .layers import (
     MLPEmbedder,
     SingleStreamBlock,
     timestep_embedding,
+    Modulation
 )
 
 @dataclass
@@ -33,6 +34,11 @@ class FluxParams:
     patch_size: int
     qkv_bias: bool
     guidance_embed: bool
+    global_modulation: bool = False
+    mlp_silu_act: bool = False
+    ops_bias: bool = True
+    default_ref_method: str = "offset"
+    ref_index_scale: float = 1.0
 
 
 class Flux(nn.Module):
@@ -58,13 +64,17 @@ class Flux(nn.Module):
         self.hidden_size = params.hidden_size
         self.num_heads = params.num_heads
         self.pe_embedder = EmbedND(dim=pe_dim, theta=params.theta, axes_dim=params.axes_dim)
-        self.img_in = operations.Linear(self.in_channels, self.hidden_size, bias=True, dtype=dtype, device=device)
-        self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations)
-        self.vector_in = MLPEmbedder(params.vec_in_dim, self.hidden_size, dtype=dtype, device=device, operations=operations)
+        self.img_in = operations.Linear(self.in_channels, self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device)
+        self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device, operations=operations)
+        if params.vec_in_dim is not None:
+            self.vector_in = MLPEmbedder(params.vec_in_dim, self.hidden_size, dtype=dtype, device=device, operations=operations)
+        else:
+            self.vector_in = None
+
         self.guidance_in = (
-            MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations) if params.guidance_embed else nn.Identity()
+            MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device, operations=operations) if params.guidance_embed else nn.Identity()
         )
-        self.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, dtype=dtype, device=device)
+        self.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device)
 
         self.double_blocks = nn.ModuleList(
             [
@@ -73,6 +83,9 @@ class Flux(nn.Module):
                     self.num_heads,
                     mlp_ratio=params.mlp_ratio,
                     qkv_bias=params.qkv_bias,
+                    modulation=params.global_modulation is False,
+                    mlp_silu_act=params.mlp_silu_act,
+                    proj_bias=params.ops_bias,
                     dtype=dtype, device=device, operations=operations
                 )
                 for _ in range(params.depth)
@@ -81,13 +94,30 @@ class Flux(nn.Module):
 
         self.single_blocks = nn.ModuleList(
             [
-                SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, dtype=dtype, device=device, operations=operations)
+                SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, modulation=params.global_modulation is False, mlp_silu_act=params.mlp_silu_act, bias=params.ops_bias, dtype=dtype, device=device, operations=operations)
                 for _ in range(params.depth_single_blocks)
             ]
         )
 
         if final_layer:
-            self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels, dtype=dtype, device=device, operations=operations)
+            self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels, bias=params.ops_bias, dtype=dtype, device=device, operations=operations)
+
+        if params.global_modulation:
+            self.double_stream_modulation_img = Modulation(
+                self.hidden_size,
+                double=True,
+                bias=False,
+                dtype=dtype, device=device, operations=operations
+            )
+            self.double_stream_modulation_txt = Modulation(
+                self.hidden_size,
+                double=True,
+                bias=False,
+                dtype=dtype, device=device, operations=operations
+            )
+            self.single_stream_modulation = Modulation(
+                self.hidden_size, double=False, bias=False, dtype=dtype, device=device, operations=operations
+            )
 
     def forward_orig(
         self,
@@ -103,9 +133,6 @@ class Flux(nn.Module):
         attn_mask: Tensor = None,
     ) -> Tensor:
 
-        if y is None:
-            y = torch.zeros((img.shape[0], self.params.vec_in_dim), device=img.device, dtype=img.dtype)
-
         patches = transformer_options.get("patches", {})
         patches_replace = transformer_options.get("patches_replace", {})
         if img.ndim != 3 or txt.ndim != 3:
@@ -118,9 +145,17 @@ class Flux(nn.Module):
             if guidance is not None:
                 vec = vec + self.guidance_in(timestep_embedding(guidance, 256).to(img.dtype))
 
-        vec = vec + self.vector_in(y[:, :self.params.vec_in_dim])
+        if self.vector_in is not None:
+            if y is None:
+                y = torch.zeros((img.shape[0], self.params.vec_in_dim), device=img.device, dtype=img.dtype)
+            vec = vec + self.vector_in(y[:, :self.params.vec_in_dim])
+
         txt = self.txt_in(txt)
 
+        vec_orig = vec
+        if self.params.global_modulation:
+            vec = (self.double_stream_modulation_img(vec_orig), self.double_stream_modulation_txt(vec_orig))
+
         if "post_input" in patches:
             for p in patches["post_input"]:
                 out = p({"img": img, "txt": txt, "img_ids": img_ids, "txt_ids": txt_ids})
@@ -177,6 +212,9 @@ class Flux(nn.Module):
 
         img = torch.cat((txt, img), 1)
 
+        if self.params.global_modulation:
+            vec, _ = self.single_stream_modulation(vec_orig)
+
         for i, block in enumerate(self.single_blocks):
             if ("single_block", i) in blocks_replace:
                 def block_wrap(args):
@@ -207,7 +245,7 @@ class Flux(nn.Module):
 
         img = img[:, txt.shape[1] :, ...]
 
-        img = self.final_layer(img, vec)  # (N, T, patch_size ** 2 * out_channels)
+        img = self.final_layer(img, vec_orig)  # (N, T, patch_size ** 2 * out_channels)
         return img
 
     def process_img(self, x, index=0, h_offset=0, w_offset=0, transformer_options={}):
@@ -234,10 +272,10 @@ class Flux(nn.Module):
             h_offset += rope_options.get("shift_y", 0.0)
             w_offset += rope_options.get("shift_x", 0.0)
 
-        img_ids = torch.zeros((steps_h, steps_w, 3), device=x.device, dtype=x.dtype)
+        img_ids = torch.zeros((steps_h, steps_w, len(self.params.axes_dim)), device=x.device, dtype=torch.float32)
         img_ids[:, :, 0] = img_ids[:, :, 1] + index
-        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=steps_h, device=x.device, dtype=x.dtype).unsqueeze(1)
-        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=steps_w, device=x.device, dtype=x.dtype).unsqueeze(0)
+        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=steps_h, device=x.device, dtype=torch.float32).unsqueeze(1)
+        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=steps_w, device=x.device, dtype=torch.float32).unsqueeze(0)
         return img, repeat(img_ids, "h w c -> b (h w) c", b=bs)
 
     def forward(self, x, timestep, context, y=None, guidance=None, ref_latents=None, control=None, transformer_options={}, **kwargs):
@@ -259,10 +297,10 @@ class Flux(nn.Module):
             h = 0
             w = 0
             index = 0
-            ref_latents_method = kwargs.get("ref_latents_method", "offset")
+            ref_latents_method = kwargs.get("ref_latents_method", self.params.default_ref_method)
             for ref in ref_latents:
                 if ref_latents_method == "index":
-                    index += 1
+                    index += self.params.ref_index_scale
                     h_offset = 0
                     w_offset = 0
                 elif ref_latents_method == "uxo":
@@ -286,7 +324,11 @@ class Flux(nn.Module):
                 img = torch.cat([img, kontext], dim=1)
                 img_ids = torch.cat([img_ids, kontext_ids], dim=1)
 
-        txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
+        txt_ids = torch.zeros((bs, context.shape[1], len(self.params.axes_dim)), device=x.device, dtype=torch.float32)
+
+        if len(self.params.axes_dim) == 4: # Flux 2
+            txt_ids[:, :, 3] = torch.linspace(0, context.shape[1] - 1, steps=context.shape[1], device=x.device, dtype=torch.float32)
+
         out = self.forward_orig(img, img_ids, context, txt_ids, timestep, y, guidance, control, transformer_options, attn_mask=kwargs.get("attention_mask", None))
         out = out[:, :img_tokens]
-        return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=2, pw=2)[:,:,:h_orig,:w_orig]
+        return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=self.patch_size, pw=self.patch_size)[:,:,:h_orig,:w_orig]

comfy/ldm/hunyuan_video/model.py

@@ -6,7 +6,6 @@ import comfy.ldm.flux.layers
 import comfy.ldm.modules.diffusionmodules.mmdit
 from comfy.ldm.modules.attention import optimized_attention
 
-
 from dataclasses import dataclass
 from einops import repeat
 
@@ -42,6 +41,8 @@ class HunyuanVideoParams:
     guidance_embed: bool
     byt5: bool
     meanflow: bool
+    use_cond_type_embedding: bool
+    vision_in_dim: int
 
 
 class SelfAttentionRef(nn.Module):
@@ -157,7 +158,10 @@ class TokenRefiner(nn.Module):
         t = self.t_embedder(timestep_embedding(timesteps, 256, time_factor=1.0).to(x.dtype))
         # m = mask.float().unsqueeze(-1)
         # c = (x.float() * m).sum(dim=1) / m.sum(dim=1) #TODO: the following works when the x.shape is the same length as the tokens but might break otherwise
-        c = x.sum(dim=1) / x.shape[1]
+        if x.dtype == torch.float16:
+            c = x.float().sum(dim=1) / x.shape[1]
+        else:
+            c = x.sum(dim=1) / x.shape[1]
 
         c = t + self.c_embedder(c.to(x.dtype))
         x = self.input_embedder(x)
@@ -196,11 +200,15 @@ class HunyuanVideo(nn.Module):
     def __init__(self, image_model=None, final_layer=True, dtype=None, device=None, operations=None, **kwargs):
         super().__init__()
         self.dtype = dtype
+        operation_settings = {"operations": operations, "device": device, "dtype": dtype}
+
         params = HunyuanVideoParams(**kwargs)
         self.params = params
         self.patch_size = params.patch_size
         self.in_channels = params.in_channels
         self.out_channels = params.out_channels
+        self.use_cond_type_embedding = params.use_cond_type_embedding
+        self.vision_in_dim = params.vision_in_dim
         if params.hidden_size % params.num_heads != 0:
             raise ValueError(
                 f"Hidden size {params.hidden_size} must be divisible by num_heads {params.num_heads}"
@@ -266,6 +274,18 @@ class HunyuanVideo(nn.Module):
         if final_layer:
             self.final_layer = LastLayer(self.hidden_size, self.patch_size[-1], self.out_channels, dtype=dtype, device=device, operations=operations)
 
+        # HunyuanVideo 1.5 specific modules
+        if self.vision_in_dim is not None:
+            from comfy.ldm.wan.model import MLPProj
+            self.vision_in = MLPProj(in_dim=self.vision_in_dim, out_dim=self.hidden_size, operation_settings=operation_settings)
+        else:
+            self.vision_in = None
+        if self.use_cond_type_embedding:
+            # 0: text_encoder feature 1: byt5 feature 2: vision_encoder feature
+            self.cond_type_embedding = nn.Embedding(3, self.hidden_size)
+        else:
+            self.cond_type_embedding = None
+
     def forward_orig(
         self,
         img: Tensor,
@@ -276,6 +296,7 @@ class HunyuanVideo(nn.Module):
         timesteps: Tensor,
         y: Tensor = None,
         txt_byt5=None,
+        clip_fea=None,
         guidance: Tensor = None,
         guiding_frame_index=None,
         ref_latent=None,
@@ -331,12 +352,31 @@ class HunyuanVideo(nn.Module):
 
         txt = self.txt_in(txt, timesteps, txt_mask, transformer_options=transformer_options)
 
+        if self.cond_type_embedding is not None:
+            self.cond_type_embedding.to(txt.device)
+            cond_emb = self.cond_type_embedding(torch.zeros_like(txt[:, :, 0], device=txt.device, dtype=torch.long))
+            txt = txt + cond_emb.to(txt.dtype)
+
         if self.byt5_in is not None and txt_byt5 is not None:
             txt_byt5 = self.byt5_in(txt_byt5)
+            if self.cond_type_embedding is not None:
+                cond_emb = self.cond_type_embedding(torch.ones_like(txt_byt5[:, :, 0], device=txt_byt5.device, dtype=torch.long))
+                txt_byt5 = txt_byt5 + cond_emb.to(txt_byt5.dtype)
+                txt = torch.cat((txt_byt5, txt), dim=1) # byt5 first for HunyuanVideo1.5
+            else:
+                txt = torch.cat((txt, txt_byt5), dim=1)
             txt_byt5_ids = torch.zeros((txt_ids.shape[0], txt_byt5.shape[1], txt_ids.shape[-1]), device=txt_ids.device, dtype=txt_ids.dtype)
-            txt = torch.cat((txt, txt_byt5), dim=1)
             txt_ids = torch.cat((txt_ids, txt_byt5_ids), dim=1)
 
+        if clip_fea is not None:
+            txt_vision_states = self.vision_in(clip_fea)
+            if self.cond_type_embedding is not None:
+                cond_emb = self.cond_type_embedding(2 * torch.ones_like(txt_vision_states[:, :, 0], dtype=torch.long, device=txt_vision_states.device))
+                txt_vision_states = txt_vision_states + cond_emb
+            txt = torch.cat((txt_vision_states.to(txt.dtype), txt), dim=1)
+            extra_txt_ids = torch.zeros((txt_ids.shape[0], txt_vision_states.shape[1], txt_ids.shape[-1]), device=txt_ids.device, dtype=txt_ids.dtype)
+            txt_ids = torch.cat((txt_ids, extra_txt_ids), dim=1)
+
         ids = torch.cat((img_ids, txt_ids), dim=1)
         pe = self.pe_embedder(ids)
 
@@ -349,7 +389,10 @@ class HunyuanVideo(nn.Module):
             attn_mask = None
 
         blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.double_blocks)
+        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.double_blocks):
+            transformer_options["block_index"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
@@ -371,7 +414,10 @@ class HunyuanVideo(nn.Module):
 
         img = torch.cat((img, txt), 1)
 
+        transformer_options["total_blocks"] = len(self.single_blocks)
+        transformer_options["block_type"] = "single"
         for i, block in enumerate(self.single_blocks):
+            transformer_options["block_index"] = i
             if ("single_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
@@ -430,14 +476,14 @@ class HunyuanVideo(nn.Module):
         img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0)
         return repeat(img_ids, "h w c -> b (h w) c", b=bs)
 
-    def forward(self, x, timestep, context, y=None, txt_byt5=None, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, disable_time_r=False, control=None, transformer_options={}, **kwargs):
+    def forward(self, x, timestep, context, y=None, txt_byt5=None, clip_fea=None, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, disable_time_r=False, control=None, transformer_options={}, **kwargs):
         return comfy.patcher_extension.WrapperExecutor.new_class_executor(
             self._forward,
             self,
             comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
-        ).execute(x, timestep, context, y, txt_byt5, guidance, attention_mask, guiding_frame_index, ref_latent, disable_time_r, control, transformer_options, **kwargs)
+        ).execute(x, timestep, context, y, txt_byt5, clip_fea, guidance, attention_mask, guiding_frame_index, ref_latent, disable_time_r, control, transformer_options, **kwargs)
 
-    def _forward(self, x, timestep, context, y=None, txt_byt5=None, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, disable_time_r=False, control=None, transformer_options={}, **kwargs):
+    def _forward(self, x, timestep, context, y=None, txt_byt5=None, clip_fea=None, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, disable_time_r=False, control=None, transformer_options={}, **kwargs):
         bs = x.shape[0]
         if len(self.patch_size) == 3:
             img_ids = self.img_ids(x)
@@ -445,5 +491,5 @@ class HunyuanVideo(nn.Module):
         else:
             img_ids = self.img_ids_2d(x)
             txt_ids = torch.zeros((bs, context.shape[1], 2), device=x.device, dtype=x.dtype)
-        out = self.forward_orig(x, img_ids, context, txt_ids, attention_mask, timestep, y, txt_byt5, guidance, guiding_frame_index, ref_latent, disable_time_r=disable_time_r, control=control, transformer_options=transformer_options)
+        out = self.forward_orig(x, img_ids, context, txt_ids, attention_mask, timestep, y, txt_byt5, clip_fea, guidance, guiding_frame_index, ref_latent, disable_time_r=disable_time_r, control=control, transformer_options=transformer_options)
         return out

comfy/ldm/hunyuan_video/upsampler.py

@@ -0,0 +1,120 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from comfy.ldm.hunyuan_video.vae_refiner import RMS_norm, ResnetBlock, VideoConv3d
+import model_management, model_patcher
+
+class SRResidualCausalBlock3D(nn.Module):
+    def __init__(self, channels: int):
+        super().__init__()
+        self.block = nn.Sequential(
+            VideoConv3d(channels, channels, kernel_size=3),
+            nn.SiLU(inplace=True),
+            VideoConv3d(channels, channels, kernel_size=3),
+            nn.SiLU(inplace=True),
+            VideoConv3d(channels, channels, kernel_size=3),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return x + self.block(x)
+
+class SRModel3DV2(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        hidden_channels: int = 64,
+        num_blocks: int = 6,
+        global_residual: bool = False,
+    ):
+        super().__init__()
+        self.in_conv = VideoConv3d(in_channels, hidden_channels, kernel_size=3)
+        self.blocks = nn.ModuleList([SRResidualCausalBlock3D(hidden_channels) for _ in range(num_blocks)])
+        self.out_conv = VideoConv3d(hidden_channels, out_channels, kernel_size=3)
+        self.global_residual = bool(global_residual)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        residual = x
+        y = self.in_conv(x)
+        for blk in self.blocks:
+            y = blk(y)
+        y = self.out_conv(y)
+        if self.global_residual and (y.shape == residual.shape):
+            y = y + residual
+        return y
+
+
+class Upsampler(nn.Module):
+    def __init__(
+        self,
+        z_channels: int,
+        out_channels: int,
+        block_out_channels: tuple[int, ...],
+        num_res_blocks: int = 2,
+    ):
+        super().__init__()
+        self.num_res_blocks = num_res_blocks
+        self.block_out_channels = block_out_channels
+        self.z_channels = z_channels
+
+        ch = block_out_channels[0]
+        self.conv_in = VideoConv3d(z_channels, ch, kernel_size=3)
+
+        self.up = nn.ModuleList()
+
+        for i, tgt in enumerate(block_out_channels):
+            stage = nn.Module()
+            stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt,
+                                                    out_channels=tgt,
+                                                    temb_channels=0,
+                                                    conv_shortcut=False,
+                                                    conv_op=VideoConv3d, norm_op=RMS_norm)
+                                        for j in range(num_res_blocks + 1)])
+            ch = tgt
+            self.up.append(stage)
+
+        self.norm_out = RMS_norm(ch)
+        self.conv_out = VideoConv3d(ch, out_channels, kernel_size=3)
+
+    def forward(self, z):
+        """
+        Args:
+            z: (B, C, T, H, W)
+            target_shape: (H, W)
+        """
+        # z to block_in
+        repeats = self.block_out_channels[0] // (self.z_channels)
+        x = self.conv_in(z) + z.repeat_interleave(repeats=repeats, dim=1)
+
+        # upsampling
+        for stage in self.up:
+            for blk in stage.block:
+                x = blk(x)
+
+        out = self.conv_out(F.silu(self.norm_out(x)))
+        return out
+
+UPSAMPLERS = {
+    "720p": SRModel3DV2,
+    "1080p": Upsampler,
+}
+
+class HunyuanVideo15SRModel():
+    def __init__(self, model_type, config):
+        self.load_device = model_management.vae_device()
+        offload_device = model_management.vae_offload_device()
+        self.dtype = model_management.vae_dtype(self.load_device)
+        self.model_class = UPSAMPLERS.get(model_type)
+        self.model = self.model_class(**config).eval()
+
+        self.patcher = model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
+
+    def load_sd(self, sd):
+        return self.model.load_state_dict(sd, strict=True)
+
+    def get_sd(self):
+        return self.model.state_dict()
+
+    def resample_latent(self, latent):
+        model_management.load_model_gpu(self.patcher)
+        return self.model(latent.to(self.load_device))

comfy/ldm/hunyuan_video/vae_refiner.py

@@ -4,8 +4,40 @@ import torch.nn.functional as F
 from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, AttnBlock, VideoConv3d, Normalize
 import comfy.ops
 import comfy.ldm.models.autoencoder
+import comfy.model_management
 ops = comfy.ops.disable_weight_init
 
+class NoPadConv3d(nn.Module):
+    def __init__(self, n_channels, out_channels, kernel_size, stride=1, dilation=1, padding=0, **kwargs):
+        super().__init__()
+        self.conv = ops.Conv3d(n_channels, out_channels, kernel_size, stride=stride, dilation=dilation, **kwargs)
+
+    def forward(self, x):
+        return self.conv(x)
+
+
+def conv_carry_causal_3d(xl, op, conv_carry_in=None, conv_carry_out=None):
+
+    x = xl[0]
+    xl.clear()
+
+    if conv_carry_out is not None:
+        to_push = x[:, :, -2:, :, :].clone()
+        conv_carry_out.append(to_push)
+
+    if isinstance(op, NoPadConv3d):
+        if conv_carry_in is None:
+            x = torch.nn.functional.pad(x, (1, 1, 1, 1, 2, 0), mode = 'replicate')
+        else:
+            carry_len = conv_carry_in[0].shape[2]
+            x = torch.cat([conv_carry_in.pop(0), x], dim=2)
+            x = torch.nn.functional.pad(x, (1, 1, 1, 1, 2 - carry_len, 0), mode = 'replicate')
+
+    out = op(x)
+
+    return out
+
+
 class RMS_norm(nn.Module):
     def __init__(self, dim):
         super().__init__()
@@ -14,7 +46,7 @@ class RMS_norm(nn.Module):
         self.gamma = nn.Parameter(torch.empty(shape))
 
     def forward(self, x):
-        return F.normalize(x, dim=1) * self.scale * self.gamma
+        return F.normalize(x, dim=1) * self.scale * comfy.model_management.cast_to(self.gamma, dtype=x.dtype, device=x.device)
 
 class DnSmpl(nn.Module):
     def __init__(self, ic, oc, tds=True, refiner_vae=True, op=VideoConv3d):
@@ -27,11 +59,12 @@ class DnSmpl(nn.Module):
         self.tds = tds
         self.gs = fct * ic // oc
 
-    def forward(self, x):
+    def forward(self, x, conv_carry_in=None, conv_carry_out=None):
         r1 = 2 if self.tds else 1
-        h = self.conv(x)
+        h = conv_carry_causal_3d([x], self.conv, conv_carry_in, conv_carry_out)
+
+        if self.tds and self.refiner_vae and conv_carry_in is None:
 
-        if self.tds and self.refiner_vae:
             hf = h[:, :, :1, :, :]
             b, c, f, ht, wd = hf.shape
             hf = hf.reshape(b, c, f, ht // 2, 2, wd // 2, 2)
@@ -39,14 +72,7 @@ class DnSmpl(nn.Module):
             hf = hf.reshape(b, 2 * 2 * c, f, ht // 2, wd // 2)
             hf = torch.cat([hf, hf], dim=1)
 
-            hn = h[:, :, 1:, :, :]
-            b, c, frms, ht, wd = hn.shape
-            nf = frms // r1
-            hn = hn.reshape(b, c, nf, r1, ht // 2, 2, wd // 2, 2)
-            hn = hn.permute(0, 3, 5, 7, 1, 2, 4, 6)
-            hn = hn.reshape(b, r1 * 2 * 2 * c, nf, ht // 2, wd // 2)
-
-            h = torch.cat([hf, hn], dim=2)
+            h = h[:, :, 1:, :, :]
 
             xf = x[:, :, :1, :, :]
             b, ci, f, ht, wd = xf.shape
@@ -54,34 +80,32 @@ class DnSmpl(nn.Module):
             xf = xf.permute(0, 4, 6, 1, 2, 3, 5)
             xf = xf.reshape(b, 2 * 2 * ci, f, ht // 2, wd // 2)
             B, C, T, H, W = xf.shape
-            xf = xf.view(B, h.shape[1], self.gs // 2, T, H, W).mean(dim=2)
+            xf = xf.view(B, hf.shape[1], self.gs // 2, T, H, W).mean(dim=2)
 
-            xn = x[:, :, 1:, :, :]
-            b, ci, frms, ht, wd = xn.shape
-            nf = frms // r1
-            xn = xn.reshape(b, ci, nf, r1, ht // 2, 2, wd // 2, 2)
-            xn = xn.permute(0, 3, 5, 7, 1, 2, 4, 6)
-            xn = xn.reshape(b, r1 * 2 * 2 * ci, nf, ht // 2, wd // 2)
-            B, C, T, H, W = xn.shape
-            xn = xn.view(B, h.shape[1], self.gs, T, H, W).mean(dim=2)
-            sc = torch.cat([xf, xn], dim=2)
-        else:
-            b, c, frms, ht, wd = h.shape
+            x = x[:, :, 1:, :, :]
 
-            nf = frms // r1
-            h = h.reshape(b, c, nf, r1, ht // 2, 2, wd // 2, 2)
-            h = h.permute(0, 3, 5, 7, 1, 2, 4, 6)
-            h = h.reshape(b, r1 * 2 * 2 * c, nf, ht // 2, wd // 2)
+        if h.shape[2] == 0:
+            return hf + xf
 
-            b, ci, frms, ht, wd = x.shape
-            nf = frms // r1
-            sc = x.reshape(b, ci, nf, r1, ht // 2, 2, wd // 2, 2)
-            sc = sc.permute(0, 3, 5, 7, 1, 2, 4, 6)
-            sc = sc.reshape(b, r1 * 2 * 2 * ci, nf, ht // 2, wd // 2)
-            B, C, T, H, W = sc.shape
-            sc = sc.view(B, h.shape[1], self.gs, T, H, W).mean(dim=2)
+        b, c, frms, ht, wd = h.shape
+        nf = frms // r1
+        h = h.reshape(b, c, nf, r1, ht // 2, 2, wd // 2, 2)
+        h = h.permute(0, 3, 5, 7, 1, 2, 4, 6)
+        h = h.reshape(b, r1 * 2 * 2 * c, nf, ht // 2, wd // 2)
 
-        return h + sc
+        b, ci, frms, ht, wd = x.shape
+        nf = frms // r1
+        x = x.reshape(b, ci, nf, r1, ht // 2, 2, wd // 2, 2)
+        x = x.permute(0, 3, 5, 7, 1, 2, 4, 6)
+        x = x.reshape(b, r1 * 2 * 2 * ci, nf, ht // 2, wd // 2)
+        B, C, T, H, W = x.shape
+        x = x.view(B, h.shape[1], self.gs, T, H, W).mean(dim=2)
+
+        if self.tds and self.refiner_vae and conv_carry_in is None:
+            h = torch.cat([hf, h], dim=2)
+            x = torch.cat([xf, x], dim=2)
+
+        return h + x
 
 
 class UpSmpl(nn.Module):
@@ -94,11 +118,11 @@ class UpSmpl(nn.Module):
         self.tus = tus
         self.rp = fct * oc // ic
 
-    def forward(self, x):
+    def forward(self, x, conv_carry_in=None, conv_carry_out=None):
         r1 = 2 if self.tus else 1
-        h = self.conv(x)
+        h = conv_carry_causal_3d([x], self.conv, conv_carry_in, conv_carry_out)
 
-        if self.tus and self.refiner_vae:
+        if self.tus and self.refiner_vae and conv_carry_in is None:
             hf = h[:, :, :1, :, :]
             b, c, f, ht, wd = hf.shape
             nc = c // (2 * 2)
@@ -107,14 +131,7 @@ class UpSmpl(nn.Module):
             hf = hf.reshape(b, nc, f, ht * 2, wd * 2)
             hf = hf[:, : hf.shape[1] // 2]
 
-            hn = h[:, :, 1:, :, :]
-            b, c, frms, ht, wd = hn.shape
-            nc = c // (r1 * 2 * 2)
-            hn = hn.reshape(b, r1, 2, 2, nc, frms, ht, wd)
-            hn = hn.permute(0, 4, 5, 1, 6, 2, 7, 3)
-            hn = hn.reshape(b, nc, frms * r1, ht * 2, wd * 2)
-
-            h = torch.cat([hf, hn], dim=2)
+            h = h[:, :, 1:, :, :]
 
             xf = x[:, :, :1, :, :]
             b, ci, f, ht, wd = xf.shape
@@ -125,29 +142,43 @@ class UpSmpl(nn.Module):
             xf = xf.permute(0, 3, 4, 5, 1, 6, 2)
             xf = xf.reshape(b, nc, f, ht * 2, wd * 2)
 
-            xn = x[:, :, 1:, :, :]
-            xn = xn.repeat_interleave(repeats=self.rp, dim=1)
-            b, c, frms, ht, wd = xn.shape
-            nc = c // (r1 * 2 * 2)
-            xn = xn.reshape(b, r1, 2, 2, nc, frms, ht, wd)
-            xn = xn.permute(0, 4, 5, 1, 6, 2, 7, 3)
-            xn = xn.reshape(b, nc, frms * r1, ht * 2, wd * 2)
-            sc = torch.cat([xf, xn], dim=2)
-        else:
-            b, c, frms, ht, wd = h.shape
-            nc = c // (r1 * 2 * 2)
-            h = h.reshape(b, r1, 2, 2, nc, frms, ht, wd)
-            h = h.permute(0, 4, 5, 1, 6, 2, 7, 3)
-            h = h.reshape(b, nc, frms * r1, ht * 2, wd * 2)
+            x = x[:, :, 1:, :, :]
 
-            sc = x.repeat_interleave(repeats=self.rp, dim=1)
-            b, c, frms, ht, wd = sc.shape
-            nc = c // (r1 * 2 * 2)
-            sc = sc.reshape(b, r1, 2, 2, nc, frms, ht, wd)
-            sc = sc.permute(0, 4, 5, 1, 6, 2, 7, 3)
-            sc = sc.reshape(b, nc, frms * r1, ht * 2, wd * 2)
+        b, c, frms, ht, wd = h.shape
+        nc = c // (r1 * 2 * 2)
+        h = h.reshape(b, r1, 2, 2, nc, frms, ht, wd)
+        h = h.permute(0, 4, 5, 1, 6, 2, 7, 3)
+        h = h.reshape(b, nc, frms * r1, ht * 2, wd * 2)
 
-        return h + sc
+        x = x.repeat_interleave(repeats=self.rp, dim=1)
+        b, c, frms, ht, wd = x.shape
+        nc = c // (r1 * 2 * 2)
+        x = x.reshape(b, r1, 2, 2, nc, frms, ht, wd)
+        x = x.permute(0, 4, 5, 1, 6, 2, 7, 3)
+        x = x.reshape(b, nc, frms * r1, ht * 2, wd * 2)
+
+        if self.tus and self.refiner_vae and conv_carry_in is None:
+            h = torch.cat([hf, h], dim=2)
+            x = torch.cat([xf, x], dim=2)
+
+        return h + x
+
+class HunyuanRefinerResnetBlock(ResnetBlock):
+    def __init__(self, in_channels, out_channels, conv_op=NoPadConv3d, norm_op=RMS_norm):
+        super().__init__(in_channels=in_channels, out_channels=out_channels, temb_channels=0, conv_op=conv_op, norm_op=norm_op)
+
+    def forward(self, x, conv_carry_in=None, conv_carry_out=None):
+        h = x
+        h = [ self.swish(self.norm1(x)) ]
+        h = conv_carry_causal_3d(h, self.conv1, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out)
+
+        h = [ self.dropout(self.swish(self.norm2(h))) ]
+        h = conv_carry_causal_3d(h, self.conv2, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out)
+
+        if self.in_channels != self.out_channels:
+            x = self.nin_shortcut(x)
+
+        return x+h
 
 class Encoder(nn.Module):
     def __init__(self, in_channels, z_channels, block_out_channels, num_res_blocks,
@@ -160,7 +191,7 @@ class Encoder(nn.Module):
 
         self.refiner_vae = refiner_vae
         if self.refiner_vae:
-            conv_op = VideoConv3d
+            conv_op = NoPadConv3d
             norm_op = RMS_norm
         else:
             conv_op = ops.Conv3d
@@ -175,10 +206,9 @@ class Encoder(nn.Module):
 
         for i, tgt in enumerate(block_out_channels):
             stage = nn.Module()
-            stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt,
-                                                     out_channels=tgt,
-                                                     temb_channels=0,
-                                                     conv_op=conv_op, norm_op=norm_op)
+            stage.block = nn.ModuleList([HunyuanRefinerResnetBlock(in_channels=ch if j == 0 else tgt,
+                                                                   out_channels=tgt,
+                                                                   conv_op=conv_op, norm_op=norm_op)
                                         for j in range(num_res_blocks)])
             ch = tgt
             if i < depth:
@@ -188,9 +218,9 @@ class Encoder(nn.Module):
             self.down.append(stage)
 
         self.mid = nn.Module()
-        self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=conv_op, norm_op=norm_op)
+        self.mid.block_1 = HunyuanRefinerResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op)
         self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=norm_op)
-        self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=conv_op, norm_op=norm_op)
+        self.mid.block_2 = HunyuanRefinerResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op)
 
         self.norm_out = norm_op(ch)
         self.conv_out = conv_op(ch, z_channels << 1, 3, 1, 1)
@@ -201,31 +231,50 @@ class Encoder(nn.Module):
         if not self.refiner_vae and x.shape[2] == 1:
             x = x.expand(-1, -1, self.ffactor_temporal, -1, -1)
 
-        x = self.conv_in(x)
+        if self.refiner_vae:
+            xl = [x[:, :, :1, :, :]]
+            if x.shape[2] > self.ffactor_temporal:
+                xl += torch.split(x[:, :, 1: 1 + ((x.shape[2] - 1) // self.ffactor_temporal) * self.ffactor_temporal, :, :], self.ffactor_temporal * 2, dim=2)
+            x = xl
+        else:
+            x = [x]
+        out = []
 
-        for stage in self.down:
-            for blk in stage.block:
-                x = blk(x)
-            if hasattr(stage, 'downsample'):
-                x = stage.downsample(x)
+        conv_carry_in = None
 
-        x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(x)))
+        for i, x1 in enumerate(x):
+            conv_carry_out = []
+            if i == len(x) - 1:
+                conv_carry_out = None
+            x1 = [ x1 ]
+            x1 = conv_carry_causal_3d(x1, self.conv_in, conv_carry_in, conv_carry_out)
+
+            for stage in self.down:
+                for blk in stage.block:
+                    x1 = blk(x1, conv_carry_in, conv_carry_out)
+                if hasattr(stage, 'downsample'):
+                    x1 = stage.downsample(x1, conv_carry_in, conv_carry_out)
+
+            out.append(x1)
+            conv_carry_in = conv_carry_out
+
+        if len(out) > 1:
+            out = torch.cat(out, dim=2)
+        else:
+            out = out[0]
+
+        x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(out)))
+        del out
 
         b, c, t, h, w = x.shape
         grp = c // (self.z_channels << 1)
         skip = x.view(b, c // grp, grp, t, h, w).mean(2)
 
-        out = self.conv_out(F.silu(self.norm_out(x))) + skip
+        out = conv_carry_causal_3d([F.silu(self.norm_out(x))], self.conv_out) + skip
 
         if self.refiner_vae:
             out = self.regul(out)[0]
 
-            out = torch.cat((out[:, :, :1], out), dim=2)
-            out = out.permute(0, 2, 1, 3, 4)
-            b, f_times_2, c, h, w = out.shape
-            out = out.reshape(b, f_times_2 // 2, 2 * c, h, w)
-            out = out.permute(0, 2, 1, 3, 4).contiguous()
-
         return out
 
 class Decoder(nn.Module):
@@ -239,7 +288,7 @@ class Decoder(nn.Module):
 
         self.refiner_vae = refiner_vae
         if self.refiner_vae:
-            conv_op = VideoConv3d
+            conv_op = NoPadConv3d
             norm_op = RMS_norm
         else:
             conv_op = ops.Conv3d
@@ -249,9 +298,9 @@ class Decoder(nn.Module):
         self.conv_in = conv_op(z_channels, ch, kernel_size=3, stride=1, padding=1)
 
         self.mid = nn.Module()
-        self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=conv_op, norm_op=norm_op)
+        self.mid.block_1 = HunyuanRefinerResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op)
         self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=norm_op)
-        self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=conv_op, norm_op=norm_op)
+        self.mid.block_2 = HunyuanRefinerResnetBlock(in_channels=ch, out_channels=ch,  conv_op=conv_op, norm_op=norm_op)
 
         self.up = nn.ModuleList()
         depth = (ffactor_spatial >> 1).bit_length()
@@ -259,10 +308,9 @@ class Decoder(nn.Module):
 
         for i, tgt in enumerate(block_out_channels):
             stage = nn.Module()
-            stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt,
-                                                     out_channels=tgt,
-                                                     temb_channels=0,
-                                                     conv_op=conv_op, norm_op=norm_op)
+            stage.block = nn.ModuleList([HunyuanRefinerResnetBlock(in_channels=ch if j == 0 else tgt,
+                                                                   out_channels=tgt,
+                                                                   conv_op=conv_op, norm_op=norm_op)
                                         for j in range(num_res_blocks + 1)])
             ch = tgt
             if i < depth:
@@ -275,27 +323,41 @@ class Decoder(nn.Module):
         self.conv_out = conv_op(ch, out_channels, 3, stride=1, padding=1)
 
     def forward(self, z):
-        if self.refiner_vae:
-            z = z.permute(0, 2, 1, 3, 4)
-            b, f, c, h, w = z.shape
-            z = z.reshape(b, f, 2, c // 2, h, w)
-            z = z.permute(0, 1, 2, 3, 4, 5).reshape(b, f * 2, c // 2, h, w)
-            z = z.permute(0, 2, 1, 3, 4)
-            z = z[:, :, 1:]
-
-        x = self.conv_in(z) + z.repeat_interleave(self.block_out_channels[0] // self.z_channels, 1)
+        x = conv_carry_causal_3d([z], self.conv_in) + z.repeat_interleave(self.block_out_channels[0] // self.z_channels, 1)
         x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(x)))
 
-        for stage in self.up:
-            for blk in stage.block:
-                x = blk(x)
-            if hasattr(stage, 'upsample'):
-                x = stage.upsample(x)
+        if self.refiner_vae:
+            x = torch.split(x, 2, dim=2)
+        else:
+            x = [ x ]
+        out = []
 
-        out = self.conv_out(F.silu(self.norm_out(x)))
+        conv_carry_in = None
+
+        for i, x1 in enumerate(x):
+            conv_carry_out = []
+            if i == len(x) - 1:
+                conv_carry_out = None
+            for stage in self.up:
+                for blk in stage.block:
+                    x1 = blk(x1, conv_carry_in, conv_carry_out)
+                if hasattr(stage, 'upsample'):
+                    x1 = stage.upsample(x1, conv_carry_in, conv_carry_out)
+
+            x1 = [ F.silu(self.norm_out(x1)) ]
+            x1 = conv_carry_causal_3d(x1, self.conv_out, conv_carry_in, conv_carry_out)
+            out.append(x1)
+            conv_carry_in = conv_carry_out
+        del x
+
+        if len(out) > 1:
+            out = torch.cat(out, dim=2)
+        else:
+            out = out[0]
 
         if not self.refiner_vae:
             if z.shape[-3] == 1:
                 out = out[:, :, -1:]
 
         return out
+

comfy/ldm/lumina/model.py

@@ -11,6 +11,7 @@ import comfy.ldm.common_dit
 from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder
 from comfy.ldm.modules.attention import optimized_attention_masked
 from comfy.ldm.flux.layers import EmbedND
+from comfy.ldm.flux.math import apply_rope
 import comfy.patcher_extension
 
 
@@ -31,6 +32,7 @@ class JointAttention(nn.Module):
         n_heads: int,
         n_kv_heads: Optional[int],
         qk_norm: bool,
+        out_bias: bool = False,
         operation_settings={},
     ):
         """
@@ -59,7 +61,7 @@ class JointAttention(nn.Module):
         self.out = operation_settings.get("operations").Linear(
             n_heads * self.head_dim,
             dim,
-            bias=False,
+            bias=out_bias,
             device=operation_settings.get("device"),
             dtype=operation_settings.get("dtype"),
         )
@@ -70,35 +72,6 @@ class JointAttention(nn.Module):
         else:
             self.q_norm = self.k_norm = nn.Identity()
 
-    @staticmethod
-    def apply_rotary_emb(
-        x_in: torch.Tensor,
-        freqs_cis: torch.Tensor,
-    ) -> torch.Tensor:
-        """
-        Apply rotary embeddings to input tensors using the given frequency
-        tensor.
-
-        This function applies rotary embeddings to the given query 'xq' and
-        key 'xk' tensors using the provided frequency tensor 'freqs_cis'. The
-        input tensors are reshaped as complex numbers, and the frequency tensor
-        is reshaped for broadcasting compatibility. The resulting tensors
-        contain rotary embeddings and are returned as real tensors.
-
-        Args:
-            x_in (torch.Tensor): Query or Key tensor to apply rotary embeddings.
-            freqs_cis (torch.Tensor): Precomputed frequency tensor for complex
-                exponentials.
-
-        Returns:
-            Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor
-                and key tensor with rotary embeddings.
-        """
-
-        t_ = x_in.reshape(*x_in.shape[:-1], -1, 1, 2)
-        t_out = freqs_cis[..., 0] * t_[..., 0] + freqs_cis[..., 1] * t_[..., 1]
-        return t_out.reshape(*x_in.shape)
-
     def forward(
         self,
         x: torch.Tensor,
@@ -134,8 +107,7 @@ class JointAttention(nn.Module):
         xq = self.q_norm(xq)
         xk = self.k_norm(xk)
 
-        xq = JointAttention.apply_rotary_emb(xq, freqs_cis=freqs_cis)
-        xk = JointAttention.apply_rotary_emb(xk, freqs_cis=freqs_cis)
+        xq, xk = apply_rope(xq, xk, freqs_cis)
 
         n_rep = self.n_local_heads // self.n_local_kv_heads
         if n_rep >= 1:
@@ -215,6 +187,8 @@ class JointTransformerBlock(nn.Module):
         norm_eps: float,
         qk_norm: bool,
         modulation=True,
+        z_image_modulation=False,
+        attn_out_bias=False,
         operation_settings={},
     ) -> None:
         """
@@ -235,10 +209,10 @@ class JointTransformerBlock(nn.Module):
         super().__init__()
         self.dim = dim
         self.head_dim = dim // n_heads
-        self.attention = JointAttention(dim, n_heads, n_kv_heads, qk_norm, operation_settings=operation_settings)
+        self.attention = JointAttention(dim, n_heads, n_kv_heads, qk_norm, out_bias=attn_out_bias, operation_settings=operation_settings)
         self.feed_forward = FeedForward(
             dim=dim,
-            hidden_dim=4 * dim,
+            hidden_dim=dim,
             multiple_of=multiple_of,
             ffn_dim_multiplier=ffn_dim_multiplier,
             operation_settings=operation_settings,
@@ -252,16 +226,27 @@ class JointTransformerBlock(nn.Module):
 
         self.modulation = modulation
         if modulation:
-            self.adaLN_modulation = nn.Sequential(
-                nn.SiLU(),
-                operation_settings.get("operations").Linear(
-                    min(dim, 1024),
-                    4 * dim,
-                    bias=True,
-                    device=operation_settings.get("device"),
-                    dtype=operation_settings.get("dtype"),
-                ),
-            )
+            if z_image_modulation:
+                self.adaLN_modulation = nn.Sequential(
+                    operation_settings.get("operations").Linear(
+                        min(dim, 256),
+                        4 * dim,
+                        bias=True,
+                        device=operation_settings.get("device"),
+                        dtype=operation_settings.get("dtype"),
+                    ),
+                )
+            else:
+                self.adaLN_modulation = nn.Sequential(
+                    nn.SiLU(),
+                    operation_settings.get("operations").Linear(
+                        min(dim, 1024),
+                        4 * dim,
+                        bias=True,
+                        device=operation_settings.get("device"),
+                        dtype=operation_settings.get("dtype"),
+                    ),
+                )
 
     def forward(
         self,
@@ -323,7 +308,7 @@ class FinalLayer(nn.Module):
     The final layer of NextDiT.
     """
 
-    def __init__(self, hidden_size, patch_size, out_channels, operation_settings={}):
+    def __init__(self, hidden_size, patch_size, out_channels, z_image_modulation=False, operation_settings={}):
         super().__init__()
         self.norm_final = operation_settings.get("operations").LayerNorm(
             hidden_size,
@@ -340,10 +325,15 @@ class FinalLayer(nn.Module):
             dtype=operation_settings.get("dtype"),
         )
 
+        if z_image_modulation:
+            min_mod = 256
+        else:
+            min_mod = 1024
+
         self.adaLN_modulation = nn.Sequential(
             nn.SiLU(),
             operation_settings.get("operations").Linear(
-                min(hidden_size, 1024),
+                min(hidden_size, min_mod),
                 hidden_size,
                 bias=True,
                 device=operation_settings.get("device"),
@@ -373,12 +363,16 @@ class NextDiT(nn.Module):
         n_heads: int = 32,
         n_kv_heads: Optional[int] = None,
         multiple_of: int = 256,
-        ffn_dim_multiplier: Optional[float] = None,
+        ffn_dim_multiplier: float = 4.0,
         norm_eps: float = 1e-5,
         qk_norm: bool = False,
         cap_feat_dim: int = 5120,
         axes_dims: List[int] = (16, 56, 56),
         axes_lens: List[int] = (1, 512, 512),
+        rope_theta=10000.0,
+        z_image_modulation=False,
+        time_scale=1.0,
+        pad_tokens_multiple=None,
         image_model=None,
         device=None,
         dtype=None,
@@ -390,6 +384,8 @@ class NextDiT(nn.Module):
         self.in_channels = in_channels
         self.out_channels = in_channels
         self.patch_size = patch_size
+        self.time_scale = time_scale
+        self.pad_tokens_multiple = pad_tokens_multiple
 
         self.x_embedder = operation_settings.get("operations").Linear(
             in_features=patch_size * patch_size * in_channels,
@@ -411,6 +407,7 @@ class NextDiT(nn.Module):
                     norm_eps,
                     qk_norm,
                     modulation=True,
+                    z_image_modulation=z_image_modulation,
                     operation_settings=operation_settings,
                 )
                 for layer_id in range(n_refiner_layers)
@@ -434,7 +431,7 @@ class NextDiT(nn.Module):
             ]
         )
 
-        self.t_embedder = TimestepEmbedder(min(dim, 1024), **operation_settings)
+        self.t_embedder = TimestepEmbedder(min(dim, 1024), output_size=256 if z_image_modulation else None, **operation_settings)
         self.cap_embedder = nn.Sequential(
             operation_settings.get("operations").RMSNorm(cap_feat_dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")),
             operation_settings.get("operations").Linear(
@@ -457,18 +454,24 @@ class NextDiT(nn.Module):
                     ffn_dim_multiplier,
                     norm_eps,
                     qk_norm,
+                    z_image_modulation=z_image_modulation,
+                    attn_out_bias=False,
                     operation_settings=operation_settings,
                 )
                 for layer_id in range(n_layers)
             ]
         )
         self.norm_final = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
-        self.final_layer = FinalLayer(dim, patch_size, self.out_channels, operation_settings=operation_settings)
+        self.final_layer = FinalLayer(dim, patch_size, self.out_channels, z_image_modulation=z_image_modulation, operation_settings=operation_settings)
+
+        if self.pad_tokens_multiple is not None:
+            self.x_pad_token = nn.Parameter(torch.empty((1, dim), device=device, dtype=dtype))
+            self.cap_pad_token = nn.Parameter(torch.empty((1, dim), device=device, dtype=dtype))
 
         assert (dim // n_heads) == sum(axes_dims)
         self.axes_dims = axes_dims
         self.axes_lens = axes_lens
-        self.rope_embedder = EmbedND(dim=dim // n_heads, theta=10000.0, axes_dim=axes_dims)
+        self.rope_embedder = EmbedND(dim=dim // n_heads, theta=rope_theta, axes_dim=axes_dims)
         self.dim = dim
         self.n_heads = n_heads
 
@@ -503,108 +506,42 @@ class NextDiT(nn.Module):
         bsz = len(x)
         pH = pW = self.patch_size
         device = x[0].device
-        dtype = x[0].dtype
 
-        if cap_mask is not None:
-            l_effective_cap_len = cap_mask.sum(dim=1).tolist()
-        else:
-            l_effective_cap_len = [num_tokens] * bsz
+        if self.pad_tokens_multiple is not None:
+            pad_extra = (-cap_feats.shape[1]) % self.pad_tokens_multiple
+            cap_feats = torch.cat((cap_feats, self.cap_pad_token.to(device=cap_feats.device, dtype=cap_feats.dtype).unsqueeze(0).repeat(cap_feats.shape[0], pad_extra, 1)), dim=1)
 
-        if cap_mask is not None and not torch.is_floating_point(cap_mask):
-            cap_mask = (cap_mask - 1).to(dtype) * torch.finfo(dtype).max
+        cap_pos_ids = torch.zeros(bsz, cap_feats.shape[1], 3, dtype=torch.float32, device=device)
+        cap_pos_ids[:, :, 0] = torch.arange(cap_feats.shape[1], dtype=torch.float32, device=device) + 1.0
 
-        img_sizes = [(img.size(1), img.size(2)) for img in x]
-        l_effective_img_len = [(H // pH) * (W // pW) for (H, W) in img_sizes]
+        B, C, H, W = x.shape
+        x = self.x_embedder(x.view(B, C, H // pH, pH, W // pW, pW).permute(0, 2, 4, 3, 5, 1).flatten(3).flatten(1, 2))
 
-        max_seq_len = max(
-            (cap_len+img_len for cap_len, img_len in zip(l_effective_cap_len, l_effective_img_len))
-        )
-        max_cap_len = max(l_effective_cap_len)
-        max_img_len = max(l_effective_img_len)
+        H_tokens, W_tokens = H // pH, W // pW
+        x_pos_ids = torch.zeros((bsz, x.shape[1], 3), dtype=torch.float32, device=device)
+        x_pos_ids[:, :, 0] = cap_feats.shape[1] + 1
+        x_pos_ids[:, :, 1] = torch.arange(H_tokens, dtype=torch.float32, device=device).view(-1, 1).repeat(1, W_tokens).flatten()
+        x_pos_ids[:, :, 2] = torch.arange(W_tokens, dtype=torch.float32, device=device).view(1, -1).repeat(H_tokens, 1).flatten()
 
-        position_ids = torch.zeros(bsz, max_seq_len, 3, dtype=torch.float32, device=device)
+        if self.pad_tokens_multiple is not None:
+            pad_extra = (-x.shape[1]) % self.pad_tokens_multiple
+            x = torch.cat((x, self.x_pad_token.to(device=x.device, dtype=x.dtype).unsqueeze(0).repeat(x.shape[0], pad_extra, 1)), dim=1)
+            x_pos_ids = torch.nn.functional.pad(x_pos_ids, (0, 0, 0, pad_extra))
 
-        for i in range(bsz):
-            cap_len = l_effective_cap_len[i]
-            img_len = l_effective_img_len[i]
-            H, W = img_sizes[i]
-            H_tokens, W_tokens = H // pH, W // pW
-            assert H_tokens * W_tokens == img_len
-
-            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.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
-
-        freqs_cis = self.rope_embedder(position_ids).movedim(1, 2).to(dtype)
-
-        # build freqs_cis for cap and image individually
-        cap_freqs_cis_shape = list(freqs_cis.shape)
-        # cap_freqs_cis_shape[1] = max_cap_len
-        cap_freqs_cis_shape[1] = cap_feats.shape[1]
-        cap_freqs_cis = torch.zeros(*cap_freqs_cis_shape, device=device, dtype=freqs_cis.dtype)
-
-        img_freqs_cis_shape = list(freqs_cis.shape)
-        img_freqs_cis_shape[1] = max_img_len
-        img_freqs_cis = torch.zeros(*img_freqs_cis_shape, device=device, dtype=freqs_cis.dtype)
-
-        for i in range(bsz):
-            cap_len = l_effective_cap_len[i]
-            img_len = l_effective_img_len[i]
-            cap_freqs_cis[i, :cap_len] = freqs_cis[i, :cap_len]
-            img_freqs_cis[i, :img_len] = freqs_cis[i, cap_len:cap_len+img_len]
+        freqs_cis = self.rope_embedder(torch.cat((cap_pos_ids, x_pos_ids), dim=1)).movedim(1, 2)
 
         # refine context
         for layer in self.context_refiner:
-            cap_feats = layer(cap_feats, cap_mask, cap_freqs_cis, transformer_options=transformer_options)
+            cap_feats = layer(cap_feats, cap_mask, freqs_cis[:, :cap_pos_ids.shape[1]], transformer_options=transformer_options)
 
-        # refine image
-        flat_x = []
-        for i in range(bsz):
-            img = x[i]
-            C, H, W = img.size()
-            img = img.view(C, H // pH, pH, W // pW, pW).permute(1, 3, 2, 4, 0).flatten(2).flatten(0, 1)
-            flat_x.append(img)
-        x = flat_x
-        padded_img_embed = torch.zeros(bsz, max_img_len, x[0].shape[-1], device=device, dtype=x[0].dtype)
-        padded_img_mask = torch.zeros(bsz, max_img_len, dtype=dtype, device=device)
-        for i in range(bsz):
-            padded_img_embed[i, :l_effective_img_len[i]] = x[i]
-            padded_img_mask[i, l_effective_img_len[i]:] = -torch.finfo(dtype).max
-
-        padded_img_embed = self.x_embedder(padded_img_embed)
-        padded_img_mask = padded_img_mask.unsqueeze(1)
+        padded_img_mask = None
         for layer in self.noise_refiner:
-            padded_img_embed = layer(padded_img_embed, padded_img_mask, img_freqs_cis, t, transformer_options=transformer_options)
-
-        if cap_mask is not None:
-            mask = torch.zeros(bsz, max_seq_len, dtype=dtype, device=device)
-            mask[:, :max_cap_len] = cap_mask[:, :max_cap_len]
-        else:
-            mask = None
-
-        padded_full_embed = torch.zeros(bsz, max_seq_len, self.dim, device=device, dtype=x[0].dtype)
-        for i in range(bsz):
-            cap_len = l_effective_cap_len[i]
-            img_len = l_effective_img_len[i]
-
-            padded_full_embed[i, :cap_len] = cap_feats[i, :cap_len]
-            padded_full_embed[i, cap_len:cap_len+img_len] = padded_img_embed[i, :img_len]
+            x = layer(x, padded_img_mask, freqs_cis[:, cap_pos_ids.shape[1]:], t, transformer_options=transformer_options)
 
+        padded_full_embed = torch.cat((cap_feats, x), dim=1)
+        mask = None
+        img_sizes = [(H, W)] * bsz
+        l_effective_cap_len = [cap_feats.shape[1]] * bsz
         return padded_full_embed, mask, img_sizes, l_effective_cap_len, freqs_cis
 
     def forward(self, x, timesteps, context, num_tokens, attention_mask=None, **kwargs):
@@ -627,7 +564,7 @@ class NextDiT(nn.Module):
         y: (N,) tensor of text tokens/features
         """
 
-        t = self.t_embedder(t, dtype=x.dtype)  # (N, D)
+        t = self.t_embedder(t * self.time_scale, dtype=x.dtype)  # (N, D)
         adaln_input = t
 
         cap_feats = self.cap_embedder(cap_feats)  # (N, L, D)  # todo check if able to batchify w.o. redundant compute

comfy/ldm/models/autoencoder.py

@@ -9,6 +9,8 @@ from comfy.ldm.modules.distributions.distributions import DiagonalGaussianDistri
 from comfy.ldm.util import get_obj_from_str, instantiate_from_config
 from comfy.ldm.modules.ema import LitEma
 import comfy.ops
+from einops import rearrange
+import comfy.model_management
 
 class DiagonalGaussianRegularizer(torch.nn.Module):
     def __init__(self, sample: bool = False):
@@ -179,6 +181,21 @@ class AutoencodingEngineLegacy(AutoencodingEngine):
         self.post_quant_conv = conv_op(embed_dim, ddconfig["z_channels"], 1)
         self.embed_dim = embed_dim
 
+        if ddconfig.get("batch_norm_latent", False):
+            self.bn_eps = 1e-4
+            self.bn_momentum = 0.1
+            self.ps = [2, 2]
+            self.bn = torch.nn.BatchNorm2d(math.prod(self.ps) * ddconfig["z_channels"],
+                                           eps=self.bn_eps,
+                                           momentum=self.bn_momentum,
+                                           affine=False,
+                                           track_running_stats=True,
+                                           )
+            self.bn.eval()
+        else:
+            self.bn = None
+
+
     def get_autoencoder_params(self) -> list:
         params = super().get_autoencoder_params()
         return params
@@ -201,11 +218,36 @@ class AutoencodingEngineLegacy(AutoencodingEngine):
             z = torch.cat(z, 0)
 
         z, reg_log = self.regularization(z)
+
+        if self.bn is not None:
+            z = rearrange(z,
+                          "... c (i pi) (j pj)  -> ... (c pi pj) i j",
+                          pi=self.ps[0],
+                          pj=self.ps[1],
+                          )
+
+            z = torch.nn.functional.batch_norm(z,
+                                               comfy.model_management.cast_to(self.bn.running_mean, dtype=z.dtype, device=z.device),
+                                               comfy.model_management.cast_to(self.bn.running_var, dtype=z.dtype, device=z.device),
+                                               momentum=self.bn_momentum,
+                                               eps=self.bn_eps)
+
         if return_reg_log:
             return z, reg_log
         return z
 
     def decode(self, z: torch.Tensor, **decoder_kwargs) -> torch.Tensor:
+        if self.bn is not None:
+            s = torch.sqrt(comfy.model_management.cast_to(self.bn.running_var.view(1, -1, 1, 1), dtype=z.dtype, device=z.device) + self.bn_eps)
+            m = comfy.model_management.cast_to(self.bn.running_mean.view(1, -1, 1, 1), dtype=z.dtype, device=z.device)
+            z = z * s + m
+            z = rearrange(
+                z,
+                "... (c pi pj) i j -> ... c (i pi) (j pj)",
+                pi=self.ps[0],
+                pj=self.ps[1],
+            )
+
         if self.max_batch_size is None:
             dec = self.post_quant_conv(z)
             dec = self.decoder(dec, **decoder_kwargs)

comfy/ldm/modules/diffusionmodules/mmdit.py

@@ -211,12 +211,14 @@ class TimestepEmbedder(nn.Module):
     Embeds scalar timesteps into vector representations.
     """
 
-    def __init__(self, hidden_size, frequency_embedding_size=256, dtype=None, device=None, operations=None):
+    def __init__(self, hidden_size, frequency_embedding_size=256, output_size=None, dtype=None, device=None, operations=None):
         super().__init__()
+        if output_size is None:
+            output_size = hidden_size
         self.mlp = nn.Sequential(
             operations.Linear(frequency_embedding_size, hidden_size, bias=True, dtype=dtype, device=device),
             nn.SiLU(),
-            operations.Linear(hidden_size, hidden_size, bias=True, dtype=dtype, device=device),
+            operations.Linear(hidden_size, output_size, bias=True, dtype=dtype, device=device),
         )
         self.frequency_embedding_size = frequency_embedding_size
 

comfy/ldm/qwen_image/model.py

@@ -439,7 +439,10 @@ class QwenImageTransformer2DModel(nn.Module):
         patches = transformer_options.get("patches", {})
         blocks_replace = patches_replace.get("dit", {})
 
+        transformer_options["total_blocks"] = len(self.transformer_blocks)
+        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.transformer_blocks):
+            transformer_options["block_index"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}

comfy/model_base.py

@@ -898,12 +898,13 @@ class Flux(BaseModel):
         attention_mask = kwargs.get("attention_mask", None)
         if attention_mask is not None:
             shape = kwargs["noise"].shape
-            mask_ref_size = kwargs["attention_mask_img_shape"]
-            # the model will pad to the patch size, and then divide
-            # essentially dividing and rounding up
-            (h_tok, w_tok) = (math.ceil(shape[2] / self.diffusion_model.patch_size), math.ceil(shape[3] / self.diffusion_model.patch_size))
-            attention_mask = utils.upscale_dit_mask(attention_mask, mask_ref_size, (h_tok, w_tok))
-            out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
+            mask_ref_size = kwargs.get("attention_mask_img_shape", None)
+            if mask_ref_size is not None:
+                # the model will pad to the patch size, and then divide
+                # essentially dividing and rounding up
+                (h_tok, w_tok) = (math.ceil(shape[2] / self.diffusion_model.patch_size), math.ceil(shape[3] / self.diffusion_model.patch_size))
+                attention_mask = utils.upscale_dit_mask(attention_mask, mask_ref_size, (h_tok, w_tok))
+                out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
 
         guidance = kwargs.get("guidance", 3.5)
         if guidance is not None:
@@ -928,6 +929,16 @@ class Flux(BaseModel):
             out['ref_latents'] = list([1, 16, sum(map(lambda a: math.prod(a.size()), ref_latents)) // 16])
         return out
 
+class Flux2(Flux):
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            target_text_len = 512
+            if cross_attn.shape[1] < target_text_len:
+                cross_attn = torch.nn.functional.pad(cross_attn, (0, 0, target_text_len - cross_attn.shape[1], 0))
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+        return out
 
 class GenmoMochi(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
@@ -1103,9 +1114,13 @@ class Lumina2(BaseModel):
             if torch.numel(attention_mask) != attention_mask.sum():
                 out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
             out['num_tokens'] = comfy.conds.CONDConstant(max(1, torch.sum(attention_mask).item()))
+
         cross_attn = kwargs.get("cross_attn", None)
         if cross_attn is not None:
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+            if 'num_tokens' not in out:
+                out['num_tokens'] = comfy.conds.CONDConstant(cross_attn.shape[1])
+
         return out
 
 class WAN21(BaseModel):
@@ -1536,3 +1551,94 @@ class HunyuanImage21Refiner(HunyuanImage21):
         out = super().extra_conds(**kwargs)
         out['disable_time_r'] = comfy.conds.CONDConstant(True)
         return out
+
+class HunyuanVideo15(HunyuanVideo):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device)
+
+    def concat_cond(self, **kwargs):
+        noise = kwargs.get("noise", None)
+        extra_channels = self.diffusion_model.img_in.proj.weight.shape[1] - noise.shape[1] - 1 #noise 32 img cond 32 + mask 1
+        if extra_channels == 0:
+            return None
+
+        image = kwargs.get("concat_latent_image", None)
+        device = kwargs["device"]
+
+        if image is None:
+            shape_image = list(noise.shape)
+            shape_image[1] = extra_channels
+            image = torch.zeros(shape_image, dtype=noise.dtype, layout=noise.layout, device=noise.device)
+        else:
+            latent_dim = self.latent_format.latent_channels
+            image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
+            for i in range(0, image.shape[1], latent_dim):
+                image[:, i: i + latent_dim] = self.process_latent_in(image[:, i: i + latent_dim])
+            image = utils.resize_to_batch_size(image, noise.shape[0])
+
+        mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
+        if mask is None:
+            mask = torch.zeros_like(noise)[:, :1]
+        else:
+            mask = 1.0 - mask
+            mask = utils.common_upscale(mask.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
+            if mask.shape[-3] < noise.shape[-3]:
+                mask = torch.nn.functional.pad(mask, (0, 0, 0, 0, 0, noise.shape[-3] - mask.shape[-3]), mode='constant', value=0)
+            mask = utils.resize_to_batch_size(mask, noise.shape[0])
+
+        return torch.cat((image, mask), dim=1)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        attention_mask = kwargs.get("attention_mask", None)
+        if attention_mask is not None:
+            if torch.numel(attention_mask) != attention_mask.sum():
+                out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        conditioning_byt5small = kwargs.get("conditioning_byt5small", None)
+        if conditioning_byt5small is not None:
+            out['txt_byt5'] = comfy.conds.CONDRegular(conditioning_byt5small)
+
+        guidance = kwargs.get("guidance", 6.0)
+        if guidance is not None:
+            out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([guidance]))
+
+        clip_vision_output = kwargs.get("clip_vision_output", None)
+        if clip_vision_output is not None:
+            out['clip_fea'] = comfy.conds.CONDRegular(clip_vision_output.last_hidden_state)
+
+        return out
+
+class HunyuanVideo15_SR_Distilled(HunyuanVideo15):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device)
+
+    def concat_cond(self, **kwargs):
+        noise = kwargs.get("noise", None)
+        image = kwargs.get("concat_latent_image", None)
+        noise_augmentation = kwargs.get("noise_augmentation", 0.0)
+        device = kwargs["device"]
+
+        if image is None:
+            image = torch.zeros([noise.shape[0], noise.shape[1] * 2 + 2, noise.shape[-3], noise.shape[-2], noise.shape[-1]], device=comfy.model_management.intermediate_device())
+        else:
+            image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
+            #image = self.process_latent_in(image) # scaling wasn't applied in reference code
+            image = utils.resize_to_batch_size(image, noise.shape[0])
+            lq_image_slice = slice(noise.shape[1] + 1, 2 * noise.shape[1] + 1)
+            if noise_augmentation > 0:
+                generator = torch.Generator(device="cpu")
+                generator.manual_seed(kwargs.get("seed", 0) - 10)
+                noise = torch.randn(image[:, lq_image_slice].shape, generator=generator, dtype=image.dtype, device="cpu").to(image.device)
+                image[:, lq_image_slice] = noise_augmentation * noise + min(1.0 - noise_augmentation, 0.75) * image[:, lq_image_slice]
+            else:
+                image[:, lq_image_slice] = 0.75 * image[:, lq_image_slice]
+        return image
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        out['disable_time_r'] = comfy.conds.CONDConstant(False)
+        return out

comfy/model_detection.py

@@ -186,30 +186,68 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
 
         guidance_keys = list(filter(lambda a: a.startswith("{}guidance_in.".format(key_prefix)), state_dict_keys))
         dit_config["guidance_embed"] = len(guidance_keys) > 0
+
+        # HunyuanVideo 1.5
+        if '{}cond_type_embedding.weight'.format(key_prefix) in state_dict_keys:
+            dit_config["use_cond_type_embedding"] = True
+        else:
+            dit_config["use_cond_type_embedding"] = False
+        if '{}vision_in.proj.0.weight'.format(key_prefix) in state_dict_keys:
+            dit_config["vision_in_dim"] = state_dict['{}vision_in.proj.0.weight'.format(key_prefix)].shape[0]
+        else:
+            dit_config["vision_in_dim"] = None
         return dit_config
 
     if '{}double_blocks.0.img_attn.norm.key_norm.scale'.format(key_prefix) in state_dict_keys and ('{}img_in.weight'.format(key_prefix) in state_dict_keys or f"{key_prefix}distilled_guidance_layer.norms.0.scale" in state_dict_keys): #Flux, Chroma or Chroma Radiance (has no img_in.weight)
         dit_config = {}
-        dit_config["image_model"] = "flux"
+        if '{}double_stream_modulation_img.lin.weight'.format(key_prefix) in state_dict_keys:
+            dit_config["image_model"] = "flux2"
+            dit_config["axes_dim"] = [32, 32, 32, 32]
+            dit_config["num_heads"] = 48
+            dit_config["mlp_ratio"] = 3.0
+            dit_config["theta"] = 2000
+            dit_config["out_channels"] = 128
+            dit_config["global_modulation"] = True
+            dit_config["vec_in_dim"] = None
+            dit_config["mlp_silu_act"] = True
+            dit_config["qkv_bias"] = False
+            dit_config["ops_bias"] = False
+            dit_config["default_ref_method"] = "index"
+            dit_config["ref_index_scale"] = 10.0
+            patch_size = 1
+        else:
+            dit_config["image_model"] = "flux"
+            dit_config["axes_dim"] = [16, 56, 56]
+            dit_config["num_heads"] = 24
+            dit_config["mlp_ratio"] = 4.0
+            dit_config["theta"] = 10000
+            dit_config["out_channels"] = 16
+            dit_config["qkv_bias"] = True
+            patch_size = 2
+
         dit_config["in_channels"] = 16
-        patch_size = 2
+        dit_config["hidden_size"] = 3072
+        dit_config["context_in_dim"] = 4096
+
         dit_config["patch_size"] = patch_size
         in_key = "{}img_in.weight".format(key_prefix)
         if in_key in state_dict_keys:
-            dit_config["in_channels"] = state_dict[in_key].shape[1] // (patch_size * patch_size)
-        dit_config["out_channels"] = 16
+            w = state_dict[in_key]
+            dit_config["in_channels"] = w.shape[1] // (patch_size * patch_size)
+            dit_config["hidden_size"] = w.shape[0]
+
+        txt_in_key = "{}txt_in.weight".format(key_prefix)
+        if txt_in_key in state_dict_keys:
+            w = state_dict[txt_in_key]
+            dit_config["context_in_dim"] = w.shape[1]
+            dit_config["hidden_size"] = w.shape[0]
+
         vec_in_key = '{}vector_in.in_layer.weight'.format(key_prefix)
         if vec_in_key in state_dict_keys:
             dit_config["vec_in_dim"] = state_dict[vec_in_key].shape[1]
-        dit_config["context_in_dim"] = 4096
-        dit_config["hidden_size"] = 3072
-        dit_config["mlp_ratio"] = 4.0
-        dit_config["num_heads"] = 24
+
         dit_config["depth"] = count_blocks(state_dict_keys, '{}double_blocks.'.format(key_prefix) + '{}.')
         dit_config["depth_single_blocks"] = count_blocks(state_dict_keys, '{}single_blocks.'.format(key_prefix) + '{}.')
-        dit_config["axes_dim"] = [16, 56, 56]
-        dit_config["theta"] = 10000
-        dit_config["qkv_bias"] = True
         if '{}distilled_guidance_layer.0.norms.0.scale'.format(key_prefix) in state_dict_keys or '{}distilled_guidance_layer.norms.0.scale'.format(key_prefix) in state_dict_keys: #Chroma
             dit_config["image_model"] = "chroma"
             dit_config["in_channels"] = 64
@@ -378,14 +416,31 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config["image_model"] = "lumina2"
         dit_config["patch_size"] = 2
         dit_config["in_channels"] = 16
-        dit_config["dim"] = 2304
-        dit_config["cap_feat_dim"] = state_dict['{}cap_embedder.1.weight'.format(key_prefix)].shape[1]
+        w = state_dict['{}cap_embedder.1.weight'.format(key_prefix)]
+        dit_config["dim"] = w.shape[0]
+        dit_config["cap_feat_dim"] = w.shape[1]
         dit_config["n_layers"] = count_blocks(state_dict_keys, '{}layers.'.format(key_prefix) + '{}.')
-        dit_config["n_heads"] = 24
-        dit_config["n_kv_heads"] = 8
         dit_config["qk_norm"] = True
-        dit_config["axes_dims"] = [32, 32, 32]
-        dit_config["axes_lens"] = [300, 512, 512]
+
+        if dit_config["dim"] == 2304: # Original Lumina 2
+            dit_config["n_heads"] = 24
+            dit_config["n_kv_heads"] = 8
+            dit_config["axes_dims"] = [32, 32, 32]
+            dit_config["axes_lens"] = [300, 512, 512]
+            dit_config["rope_theta"] = 10000.0
+            dit_config["ffn_dim_multiplier"] = 4.0
+        elif dit_config["dim"] == 3840:  # Z image
+            dit_config["n_heads"] = 30
+            dit_config["n_kv_heads"] = 30
+            dit_config["axes_dims"] = [32, 48, 48]
+            dit_config["axes_lens"] = [1536, 512, 512]
+            dit_config["rope_theta"] = 256.0
+            dit_config["ffn_dim_multiplier"] = (8.0 / 3.0)
+            dit_config["z_image_modulation"] = True
+            dit_config["time_scale"] = 1000.0
+            if '{}cap_pad_token'.format(key_prefix) in state_dict_keys:
+                dit_config["pad_tokens_multiple"] = 32
+
         return dit_config
 
     if '{}head.modulation'.format(key_prefix) in state_dict_keys:  # Wan 2.1

comfy/model_management.py

@@ -1098,13 +1098,14 @@ if not args.disable_pinned_memory:
             MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.95
         logging.info("Enabled pinned memory {}".format(MAX_PINNED_MEMORY // (1024 * 1024)))
 
+PINNING_ALLOWED_TYPES = set(["Parameter", "QuantizedTensor"])
 
 def pin_memory(tensor):
     global TOTAL_PINNED_MEMORY
     if MAX_PINNED_MEMORY <= 0:
         return False
 
-    if type(tensor) is not torch.nn.parameter.Parameter:
+    if type(tensor).__name__ not in PINNING_ALLOWED_TYPES:
         return False
 
     if not is_device_cpu(tensor.device):
@@ -1124,6 +1125,9 @@ def pin_memory(tensor):
         return False
 
     ptr = tensor.data_ptr()
+    if ptr == 0:
+        return False
+
     if torch.cuda.cudart().cudaHostRegister(ptr, size, 1) == 0:
         PINNED_MEMORY[ptr] = size
         TOTAL_PINNED_MEMORY += size

comfy/model_patcher.py

@@ -132,7 +132,7 @@ class LowVramPatch:
     def __call__(self, weight):
         intermediate_dtype = weight.dtype
         if self.convert_func is not None:
-            weight = self.convert_func(weight.to(dtype=torch.float32, copy=True), inplace=True)
+            weight = self.convert_func(weight, inplace=False)
 
         if intermediate_dtype not in [torch.float32, torch.float16, torch.bfloat16]: #intermediate_dtype has to be one that is supported in math ops
             intermediate_dtype = torch.float32
@@ -231,7 +231,6 @@ class ModelPatcher:
         self.object_patches_backup = {}
         self.weight_wrapper_patches = {}
         self.model_options = {"transformer_options":{}}
-        self.model_size()
         self.load_device = load_device
         self.offload_device = offload_device
         self.weight_inplace_update = weight_inplace_update
@@ -286,7 +285,7 @@ class ModelPatcher:
         return self.model.lowvram_patch_counter
 
     def clone(self):
-        n = self.__class__(self.model, self.load_device, self.offload_device, self.size, weight_inplace_update=self.weight_inplace_update)
+        n = self.__class__(self.model, self.load_device, self.offload_device, self.model_size(), weight_inplace_update=self.weight_inplace_update)
         n.patches = {}
         for k in self.patches:
             n.patches[k] = self.patches[k][:]

comfy/ops.py

@@ -117,6 +117,8 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of
     if weight_has_function or weight.dtype != dtype:
         with wf_context:
             weight = weight.to(dtype=dtype)
+            if isinstance(weight, QuantizedTensor):
+                weight = weight.dequantize()
             for f in s.weight_function:
                 weight = f(weight)
 
@@ -502,7 +504,7 @@ def scaled_fp8_ops(fp8_matrix_mult=False, scale_input=False, override_dtype=None
                     weight *= self.scale_weight.to(device=weight.device, dtype=weight.dtype)
                     return weight
                 else:
-                    return weight * self.scale_weight.to(device=weight.device, dtype=weight.dtype)
+                    return weight.to(dtype=torch.float32) * self.scale_weight.to(device=weight.device, dtype=torch.float32)
 
             def set_weight(self, weight, inplace_update=False, seed=None, return_weight=False, **kwargs):
                 weight = comfy.float.stochastic_rounding(weight / self.scale_weight.to(device=weight.device, dtype=weight.dtype), self.weight.dtype, seed=seed)
@@ -540,115 +542,136 @@ if CUBLAS_IS_AVAILABLE:
 # ==============================================================================
 from .quant_ops import QuantizedTensor, QUANT_ALGOS
 
-class MixedPrecisionOps(disable_weight_init):
-    _layer_quant_config = {}
-    _compute_dtype = torch.bfloat16
 
-    class Linear(torch.nn.Module, CastWeightBiasOp):
-        def __init__(
-            self,
-            in_features: int,
-            out_features: int,
-            bias: bool = True,
-            device=None,
-            dtype=None,
-        ) -> None:
-            super().__init__()
+def mixed_precision_ops(layer_quant_config={}, compute_dtype=torch.bfloat16, full_precision_mm=False):
+    class MixedPrecisionOps(manual_cast):
+        _layer_quant_config = layer_quant_config
+        _compute_dtype = compute_dtype
+        _full_precision_mm = full_precision_mm
 
-            self.factory_kwargs = {"device": device, "dtype": MixedPrecisionOps._compute_dtype}
-            # self.factory_kwargs = {"device": device, "dtype": dtype}
+        class Linear(torch.nn.Module, CastWeightBiasOp):
+            def __init__(
+                self,
+                in_features: int,
+                out_features: int,
+                bias: bool = True,
+                device=None,
+                dtype=None,
+            ) -> None:
+                super().__init__()
 
-            self.in_features = in_features
-            self.out_features = out_features
-            if bias:
-                self.bias = torch.nn.Parameter(torch.empty(out_features, **self.factory_kwargs))
-            else:
-                self.register_parameter("bias", None)
+                self.factory_kwargs = {"device": device, "dtype": MixedPrecisionOps._compute_dtype}
+                # self.factory_kwargs = {"device": device, "dtype": dtype}
 
-            self.tensor_class = None
+                self.in_features = in_features
+                self.out_features = out_features
+                if bias:
+                    self.bias = torch.nn.Parameter(torch.empty(out_features, **self.factory_kwargs))
+                else:
+                    self.register_parameter("bias", None)
 
-        def reset_parameters(self):
-            return None
+                self.tensor_class = None
+                self._full_precision_mm = MixedPrecisionOps._full_precision_mm
 
-        def _load_from_state_dict(self, state_dict, prefix, local_metadata,
-                                  strict, missing_keys, unexpected_keys, error_msgs):
+            def reset_parameters(self):
+                return None
 
-            device = self.factory_kwargs["device"]
-            layer_name = prefix.rstrip('.')
-            weight_key = f"{prefix}weight"
-            weight = state_dict.pop(weight_key, None)
-            if weight is None:
-                raise ValueError(f"Missing weight for layer {layer_name}")
+            def _load_from_state_dict(self, state_dict, prefix, local_metadata,
+                                    strict, missing_keys, unexpected_keys, error_msgs):
 
-            manually_loaded_keys = [weight_key]
+                device = self.factory_kwargs["device"]
+                layer_name = prefix.rstrip('.')
+                weight_key = f"{prefix}weight"
+                weight = state_dict.pop(weight_key, None)
+                if weight is None:
+                    raise ValueError(f"Missing weight for layer {layer_name}")
 
-            if layer_name not in MixedPrecisionOps._layer_quant_config:
-                self.weight = torch.nn.Parameter(weight.to(device=device, dtype=MixedPrecisionOps._compute_dtype), requires_grad=False)
-            else:
-                quant_format = MixedPrecisionOps._layer_quant_config[layer_name].get("format", None)
-                if quant_format is None:
-                    raise ValueError(f"Unknown quantization format for layer {layer_name}")
+                manually_loaded_keys = [weight_key]
 
-                qconfig = QUANT_ALGOS[quant_format]
-                self.layout_type = qconfig["comfy_tensor_layout"]
+                if layer_name not in MixedPrecisionOps._layer_quant_config:
+                    self.weight = torch.nn.Parameter(weight.to(device=device, dtype=MixedPrecisionOps._compute_dtype), requires_grad=False)
+                else:
+                    quant_format = MixedPrecisionOps._layer_quant_config[layer_name].get("format", None)
+                    if quant_format is None:
+                        raise ValueError(f"Unknown quantization format for layer {layer_name}")
 
-                weight_scale_key = f"{prefix}weight_scale"
-                layout_params = {
-                    'scale': state_dict.pop(weight_scale_key, None),
-                    'orig_dtype': MixedPrecisionOps._compute_dtype,
-                    'block_size': qconfig.get("group_size", None),
-                }
-                if layout_params['scale'] is not None:
-                    manually_loaded_keys.append(weight_scale_key)
+                    qconfig = QUANT_ALGOS[quant_format]
+                    self.layout_type = qconfig["comfy_tensor_layout"]
 
-                self.weight = torch.nn.Parameter(
-                    QuantizedTensor(weight.to(device=device), self.layout_type, layout_params),
-                    requires_grad=False
-                )
+                    weight_scale_key = f"{prefix}weight_scale"
+                    layout_params = {
+                        'scale': state_dict.pop(weight_scale_key, None),
+                        'orig_dtype': MixedPrecisionOps._compute_dtype,
+                        'block_size': qconfig.get("group_size", None),
+                    }
+                    if layout_params['scale'] is not None:
+                        manually_loaded_keys.append(weight_scale_key)
 
-                for param_name in qconfig["parameters"]:
-                    param_key = f"{prefix}{param_name}"
-                    _v = state_dict.pop(param_key, None)
-                    if _v is None:
-                        continue
-                    setattr(self, param_name, torch.nn.Parameter(_v.to(device=device), requires_grad=False))
-                    manually_loaded_keys.append(param_key)
+                    self.weight = torch.nn.Parameter(
+                        QuantizedTensor(weight.to(device=device), self.layout_type, layout_params),
+                        requires_grad=False
+                    )
 
-            super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs)
+                    for param_name in qconfig["parameters"]:
+                        param_key = f"{prefix}{param_name}"
+                        _v = state_dict.pop(param_key, None)
+                        if _v is None:
+                            continue
+                        setattr(self, param_name, torch.nn.Parameter(_v.to(device=device), requires_grad=False))
+                        manually_loaded_keys.append(param_key)
 
-            for key in manually_loaded_keys:
-                if key in missing_keys:
-                    missing_keys.remove(key)
+                super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs)
 
-        def _forward(self, input, weight, bias):
-            return torch.nn.functional.linear(input, weight, bias)
+                for key in manually_loaded_keys:
+                    if key in missing_keys:
+                        missing_keys.remove(key)
 
-        def forward_comfy_cast_weights(self, input):
-            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, weight, bias):
+                return torch.nn.functional.linear(input, weight, bias)
 
-        def forward(self, input, *args, **kwargs):
-            run_every_op()
+            def forward_comfy_cast_weights(self, input):
+                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
 
-            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
-                return self.forward_comfy_cast_weights(input, *args, **kwargs)
-            if (getattr(self, 'layout_type', None) is not None and
-                getattr(self, 'input_scale', None) is not None and
-                not isinstance(input, QuantizedTensor)):
-                input = QuantizedTensor.from_float(input, self.layout_type, scale=self.input_scale, dtype=self.weight.dtype)
-            return self._forward(input, self.weight, self.bias)
+            def forward(self, input, *args, **kwargs):
+                run_every_op()
 
+                if self._full_precision_mm or self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
+                    return self.forward_comfy_cast_weights(input, *args, **kwargs)
+                if (getattr(self, 'layout_type', None) is not None and
+                    getattr(self, 'input_scale', None) is not None and
+                    not isinstance(input, QuantizedTensor)):
+                    input = QuantizedTensor.from_float(input, self.layout_type, scale=self.input_scale, dtype=self.weight.dtype)
+                return self._forward(input, self.weight, self.bias)
+
+            def convert_weight(self, weight, inplace=False, **kwargs):
+                if isinstance(weight, QuantizedTensor):
+                    return weight.dequantize()
+                else:
+                    return weight
+
+            def set_weight(self, weight, inplace_update=False, seed=None, return_weight=False, **kwargs):
+                if getattr(self, 'layout_type', None) is not None:
+                    weight = QuantizedTensor.from_float(weight, self.layout_type, scale=None, dtype=self.weight.dtype, stochastic_rounding=seed, inplace_ops=True)
+                else:
+                    weight = weight.to(self.weight.dtype)
+                if return_weight:
+                    return weight
+
+                assert inplace_update is False  # TODO: eventually remove the inplace_update stuff
+                self.weight = torch.nn.Parameter(weight, requires_grad=False)
+
+    return MixedPrecisionOps
 
 def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, scaled_fp8=None, model_config=None):
-    if model_config and hasattr(model_config, 'layer_quant_config') and model_config.layer_quant_config:
-        MixedPrecisionOps._layer_quant_config = model_config.layer_quant_config
-        MixedPrecisionOps._compute_dtype = compute_dtype
-        logging.info(f"Using mixed precision operations: {len(model_config.layer_quant_config)} quantized layers")
-        return MixedPrecisionOps
+    fp8_compute = comfy.model_management.supports_fp8_compute(load_device) # TODO: if we support more ops this needs to be more granular
+
+    if model_config and hasattr(model_config, 'layer_quant_config') and model_config.layer_quant_config:
+        logging.info(f"Using mixed precision operations: {len(model_config.layer_quant_config)} quantized layers")
+        return mixed_precision_ops(model_config.layer_quant_config, compute_dtype, full_precision_mm=not fp8_compute)
 
-    fp8_compute = comfy.model_management.supports_fp8_compute(load_device)
     if scaled_fp8 is not None:
         return scaled_fp8_ops(fp8_matrix_mult=fp8_compute and fp8_optimizations, scale_input=fp8_optimizations, override_dtype=scaled_fp8)
 

comfy/quant_ops.py

@@ -1,6 +1,7 @@
 import torch
 import logging
 from typing import Tuple, Dict
+import comfy.float
 
 _LAYOUT_REGISTRY = {}
 _GENERIC_UTILS = {}
@@ -228,6 +229,14 @@ class QuantizedTensor(torch.Tensor):
         new_kwargs = dequant_arg(kwargs)
         return func(*new_args, **new_kwargs)
 
+    def data_ptr(self):
+        return self._qdata.data_ptr()
+
+    def is_pinned(self):
+        return self._qdata.is_pinned()
+
+    def is_contiguous(self):
+        return self._qdata.is_contiguous()
 
 # ==============================================================================
 # Generic Utilities (Layout-Agnostic Operations)
@@ -338,6 +347,18 @@ def generic_copy_(func, args, kwargs):
     return func(*args, **kwargs)
 
 
+@register_generic_util(torch.ops.aten.to.dtype)
+def generic_to_dtype(func, args, kwargs):
+    """Handle .to(dtype) calls - dtype conversion only."""
+    src = args[0]
+    if isinstance(src, QuantizedTensor):
+        # For dtype-only conversion, just change the orig_dtype, no real cast is needed
+        target_dtype = args[1] if len(args) > 1 else kwargs.get('dtype')
+        src._layout_params["orig_dtype"] = target_dtype
+        return src
+    return func(*args, **kwargs)
+
+
 @register_generic_util(torch.ops.aten._has_compatible_shallow_copy_type.default)
 def generic_has_compatible_shallow_copy_type(func, args, kwargs):
     return True
@@ -373,7 +394,7 @@ class TensorCoreFP8Layout(QuantizedLayout):
     - orig_dtype: Original dtype before quantization (for casting back)
     """
     @classmethod
-    def quantize(cls, tensor, scale=None, dtype=torch.float8_e4m3fn):
+    def quantize(cls, tensor, scale=None, dtype=torch.float8_e4m3fn, stochastic_rounding=0, inplace_ops=False):
         orig_dtype = tensor.dtype
 
         if scale is None:
@@ -383,17 +404,23 @@ class TensorCoreFP8Layout(QuantizedLayout):
             scale = torch.tensor(scale)
         scale = scale.to(device=tensor.device, dtype=torch.float32)
 
-        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)
+        if inplace_ops:
+            tensor *= (1.0 / scale).to(tensor.dtype)
+        else:
+            tensor = tensor * (1.0 / scale).to(tensor.dtype)
+
+        if stochastic_rounding > 0:
+            tensor = comfy.float.stochastic_rounding(tensor, dtype=dtype, seed=stochastic_rounding)
+        else:
+            lp_amax = torch.finfo(dtype).max
+            torch.clamp(tensor, min=-lp_amax, max=lp_amax, out=tensor)
+            tensor = tensor.to(dtype, memory_format=torch.contiguous_format)
 
         layout_params = {
             'scale': scale,
             'orig_dtype': orig_dtype
         }
-        return qdata, layout_params
+        return tensor, layout_params
 
     @staticmethod
     def dequantize(qdata, scale, orig_dtype, **kwargs):

comfy/sd.py

@@ -52,6 +52,7 @@ import comfy.text_encoders.ace
 import comfy.text_encoders.omnigen2
 import comfy.text_encoders.qwen_image
 import comfy.text_encoders.hunyuan_image
+import comfy.text_encoders.z_image
 
 import comfy.model_patcher
 import comfy.lora
@@ -356,7 +357,7 @@ class VAE:
 
                     self.memory_used_encode = lambda shape, dtype: (700 * shape[2] * shape[3]) * model_management.dtype_size(dtype)
                     self.memory_used_decode = lambda shape, dtype: (700 * shape[2] * shape[3] * 32 * 32) * model_management.dtype_size(dtype)
-                elif sd['decoder.conv_in.weight'].shape[1] == 32:
+                elif sd['decoder.conv_in.weight'].shape[1] == 32 and sd['decoder.conv_in.weight'].ndim == 5:
                     ddconfig = {"block_out_channels": [128, 256, 512, 1024, 1024], "in_channels": 3, "out_channels": 3, "num_res_blocks": 2, "ffactor_spatial": 16, "ffactor_temporal": 4, "downsample_match_channel": True, "upsample_match_channel": True, "refiner_vae": False}
                     self.latent_channels = ddconfig['z_channels'] = sd["decoder.conv_in.weight"].shape[1]
                     self.working_dtypes = [torch.float16, torch.bfloat16, torch.float32]
@@ -382,6 +383,17 @@ class VAE:
                         self.upscale_ratio = 4
 
                     self.latent_channels = ddconfig['z_channels'] = sd["decoder.conv_in.weight"].shape[1]
+                    if 'decoder.post_quant_conv.weight' in sd:
+                        sd = comfy.utils.state_dict_prefix_replace(sd, {"decoder.post_quant_conv.": "post_quant_conv.", "encoder.quant_conv.": "quant_conv."})
+
+                    if 'bn.running_mean' in sd:
+                        ddconfig["batch_norm_latent"] = True
+                        self.downscale_ratio *= 2
+                        self.upscale_ratio *= 2
+                        self.latent_channels *= 4
+                        old_memory_used_decode = self.memory_used_decode
+                        self.memory_used_decode = lambda shape, dtype: old_memory_used_decode(shape, dtype) *  4.0
+
                     if 'post_quant_conv.weight' in sd:
                         self.first_stage_model = AutoencoderKL(ddconfig=ddconfig, embed_dim=sd['post_quant_conv.weight'].shape[1])
                     else:
@@ -441,20 +453,20 @@ class VAE:
             elif "decoder.conv_in.conv.weight" in sd and sd['decoder.conv_in.conv.weight'].shape[1] == 32:
                 ddconfig = {"block_out_channels": [128, 256, 512, 1024, 1024], "in_channels": 3, "out_channels": 3, "num_res_blocks": 2, "ffactor_spatial": 16, "ffactor_temporal": 4, "downsample_match_channel": True, "upsample_match_channel": True}
                 ddconfig['z_channels'] = sd["decoder.conv_in.conv.weight"].shape[1]
-                self.latent_channels = 64
+                self.latent_channels = 32
                 self.upscale_ratio = (lambda a: max(0, a * 4 - 3), 16, 16)
                 self.upscale_index_formula = (4, 16, 16)
                 self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 16, 16)
                 self.downscale_index_formula = (4, 16, 16)
                 self.latent_dim = 3
-                self.not_video = True
+                self.not_video = False
                 self.working_dtypes = [torch.float16, torch.bfloat16, torch.float32]
                 self.first_stage_model = AutoencodingEngine(regularizer_config={'target': "comfy.ldm.models.autoencoder.EmptyRegularizer"},
                                                             encoder_config={'target': "comfy.ldm.hunyuan_video.vae_refiner.Encoder", 'params': ddconfig},
                                                             decoder_config={'target': "comfy.ldm.hunyuan_video.vae_refiner.Decoder", 'params': ddconfig})
 
-                self.memory_used_encode = lambda shape, dtype: (1400 * shape[-2] * shape[-1]) * model_management.dtype_size(dtype)
-                self.memory_used_decode = lambda shape, dtype: (1400 * shape[-3] * shape[-2] * shape[-1] * 16 * 16) * model_management.dtype_size(dtype)
+                self.memory_used_encode = lambda shape, dtype: (1400 * 9 * shape[-2] * shape[-1]) * model_management.dtype_size(dtype)
+                self.memory_used_decode = lambda shape, dtype: (2800 * 4 * shape[-2] * shape[-1] * 16 * 16) * model_management.dtype_size(dtype)
             elif "decoder.conv_in.conv.weight" in sd:
                 ddconfig = {'double_z': True, 'z_channels': 4, 'resolution': 256, 'in_channels': 3, 'out_ch': 3, 'ch': 128, 'ch_mult': [1, 2, 4, 4], 'num_res_blocks': 2, 'attn_resolutions': [], 'dropout': 0.0}
                 ddconfig["conv3d"] = True
@@ -911,12 +923,18 @@ class CLIPType(Enum):
     OMNIGEN2 = 17
     QWEN_IMAGE = 18
     HUNYUAN_IMAGE = 19
+    HUNYUAN_VIDEO_15 = 20
 
 
 def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}):
     clip_data = []
     for p in ckpt_paths:
-        clip_data.append(comfy.utils.load_torch_file(p, safe_load=True))
+        sd, metadata = comfy.utils.load_torch_file(p, safe_load=True, return_metadata=True)
+        if metadata is not None:
+            quant_metadata = metadata.get("_quantization_metadata", None)
+            if quant_metadata is not None:
+                sd["_quantization_metadata"] = quant_metadata
+        clip_data.append(sd)
     return load_text_encoder_state_dicts(clip_data, embedding_directory=embedding_directory, clip_type=clip_type, model_options=model_options)
 
 
@@ -934,6 +952,10 @@ class TEModel(Enum):
     QWEN25_7B = 11
     BYT5_SMALL_GLYPH = 12
     GEMMA_3_4B = 13
+    MISTRAL3_24B = 14
+    MISTRAL3_24B_PRUNED_FLUX2 = 15
+    QWEN3_4B = 16
+
 
 def detect_te_model(sd):
     if "text_model.encoder.layers.30.mlp.fc1.weight" in sd:
@@ -966,6 +988,15 @@ def detect_te_model(sd):
         if weight.shape[0] == 512:
             return TEModel.QWEN25_7B
     if "model.layers.0.post_attention_layernorm.weight" in sd:
+        if 'model.layers.0.self_attn.q_norm.weight' in sd:
+            return TEModel.QWEN3_4B
+        weight = sd['model.layers.0.post_attention_layernorm.weight']
+        if weight.shape[0] == 5120:
+            if "model.layers.39.post_attention_layernorm.weight" in sd:
+                return TEModel.MISTRAL3_24B
+            else:
+                return TEModel.MISTRAL3_24B_PRUNED_FLUX2
+
         return TEModel.LLAMA3_8
     return None
 
@@ -1080,6 +1111,13 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
             else:
                 clip_target.clip = comfy.text_encoders.qwen_image.te(**llama_detect(clip_data))
                 clip_target.tokenizer = comfy.text_encoders.qwen_image.QwenImageTokenizer
+        elif te_model == TEModel.MISTRAL3_24B or te_model == TEModel.MISTRAL3_24B_PRUNED_FLUX2:
+            clip_target.clip = comfy.text_encoders.flux.flux2_te(**llama_detect(clip_data), pruned=te_model == TEModel.MISTRAL3_24B_PRUNED_FLUX2)
+            clip_target.tokenizer = comfy.text_encoders.flux.Flux2Tokenizer
+            tokenizer_data["tekken_model"] = clip_data[0].get("tekken_model", None)
+        elif te_model == TEModel.QWEN3_4B:
+            clip_target.clip = comfy.text_encoders.z_image.te(**llama_detect(clip_data))
+            clip_target.tokenizer = comfy.text_encoders.z_image.ZImageTokenizer
         else:
             # clip_l
             if clip_type == CLIPType.SD3:
@@ -1126,6 +1164,9 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
         elif clip_type == CLIPType.HUNYUAN_IMAGE:
             clip_target.clip = comfy.text_encoders.hunyuan_image.te(**llama_detect(clip_data))
             clip_target.tokenizer = comfy.text_encoders.hunyuan_image.HunyuanImageTokenizer
+        elif clip_type == CLIPType.HUNYUAN_VIDEO_15:
+            clip_target.clip = comfy.text_encoders.hunyuan_image.te(**llama_detect(clip_data))
+            clip_target.tokenizer = comfy.text_encoders.hunyuan_video.HunyuanVideo15Tokenizer
         else:
             clip_target.clip = sdxl_clip.SDXLClipModel
             clip_target.tokenizer = sdxl_clip.SDXLTokenizer
@@ -1138,6 +1179,8 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
 
     parameters = 0
     for c in clip_data:
+        if "_quantization_metadata" in c:
+            c.pop("_quantization_metadata")
         parameters += comfy.utils.calculate_parameters(c)
         tokenizer_data, model_options = comfy.text_encoders.long_clipl.model_options_long_clip(c, tokenizer_data, model_options)
 

comfy/sd1_clip.py

@@ -90,7 +90,6 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
                  special_tokens={"start": 49406, "end": 49407, "pad": 49407}, layer_norm_hidden_state=True, enable_attention_masks=False, zero_out_masked=False,
                  return_projected_pooled=True, return_attention_masks=False, model_options={}):  # clip-vit-base-patch32
         super().__init__()
-        assert layer in self.LAYERS
 
         if textmodel_json_config is None:
             textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_clip_config.json")
@@ -109,13 +108,23 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
 
         operations = model_options.get("custom_operations", None)
         scaled_fp8 = None
+        quantization_metadata = model_options.get("quantization_metadata", None)
 
         if operations is None:
-            scaled_fp8 = model_options.get("scaled_fp8", None)
-            if scaled_fp8 is not None:
-                operations = comfy.ops.scaled_fp8_ops(fp8_matrix_mult=False, override_dtype=scaled_fp8)
+            layer_quant_config = None
+            if quantization_metadata is not None:
+                layer_quant_config = json.loads(quantization_metadata).get("layers", None)
+
+            if layer_quant_config is not None:
+                operations = comfy.ops.mixed_precision_ops(layer_quant_config, dtype, full_precision_mm=True)
+                logging.info(f"Using MixedPrecisionOps for text encoder: {len(layer_quant_config)} quantized layers")
             else:
-                operations = comfy.ops.manual_cast
+                # Fallback to scaled_fp8_ops for backward compatibility
+                scaled_fp8 = model_options.get("scaled_fp8", None)
+                if scaled_fp8 is not None:
+                    operations = comfy.ops.scaled_fp8_ops(fp8_matrix_mult=False, override_dtype=scaled_fp8)
+                else:
+                    operations = comfy.ops.manual_cast
 
         self.operations = operations
         self.transformer = model_class(config, dtype, device, self.operations)
@@ -154,7 +163,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
     def set_clip_options(self, options):
         layer_idx = options.get("layer", self.layer_idx)
         self.return_projected_pooled = options.get("projected_pooled", self.return_projected_pooled)
-        if self.layer == "all":
+        if isinstance(self.layer, list) or self.layer == "all":
             pass
         elif layer_idx is None or abs(layer_idx) > self.num_layers:
             self.layer = "last"
@@ -256,7 +265,9 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
         if self.enable_attention_masks:
             attention_mask_model = attention_mask
 
-        if self.layer == "all":
+        if isinstance(self.layer, list):
+            intermediate_output = self.layer
+        elif self.layer == "all":
             intermediate_output = "all"
         else:
             intermediate_output = self.layer_idx

comfy/supported_models.py

@@ -21,6 +21,7 @@ import comfy.text_encoders.ace
 import comfy.text_encoders.omnigen2
 import comfy.text_encoders.qwen_image
 import comfy.text_encoders.hunyuan_image
+import comfy.text_encoders.z_image
 
 from . import supported_models_base
 from . import latent_formats
@@ -741,6 +742,37 @@ class FluxSchnell(Flux):
         out = model_base.Flux(self, model_type=model_base.ModelType.FLOW, device=device)
         return out
 
+class Flux2(Flux):
+    unet_config = {
+        "image_model": "flux2",
+    }
+
+    sampling_settings = {
+        "shift": 2.02,
+    }
+
+    unet_extra_config = {}
+    latent_format = latent_formats.Flux2
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
+
+    vae_key_prefix = ["vae."]
+    text_encoder_key_prefix = ["text_encoders."]
+
+    def __init__(self, unet_config):
+        super().__init__(unet_config)
+        self.memory_usage_factor = self.memory_usage_factor * (2.0 * 2.0) * 2.36
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.Flux2(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        return None # TODO
+        pref = self.text_encoder_key_prefix[0]
+        t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}t5xxl.transformer.".format(pref))
+        return supported_models_base.ClipTarget(comfy.text_encoders.flux.FluxTokenizer, comfy.text_encoders.flux.flux_clip(**t5_detect))
+
 class GenmoMochi(supported_models_base.BASE):
     unet_config = {
         "image_model": "mochi_preview",
@@ -963,7 +995,7 @@ class Lumina2(supported_models_base.BASE):
         "shift": 6.0,
     }
 
-    memory_usage_factor = 1.2
+    memory_usage_factor = 1.4
 
     unet_extra_config = {}
     latent_format = latent_formats.Flux
@@ -982,6 +1014,24 @@ class Lumina2(supported_models_base.BASE):
         hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}gemma2_2b.transformer.".format(pref))
         return supported_models_base.ClipTarget(comfy.text_encoders.lumina2.LuminaTokenizer, comfy.text_encoders.lumina2.te(**hunyuan_detect))
 
+class ZImage(Lumina2):
+    unet_config = {
+        "image_model": "lumina2",
+        "dim": 3840,
+    }
+
+    sampling_settings = {
+        "multiplier": 1.0,
+        "shift": 3.0,
+    }
+
+    memory_usage_factor = 1.7
+
+    def clip_target(self, state_dict={}):
+        pref = self.text_encoder_key_prefix[0]
+        hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_4b.transformer.".format(pref))
+        return supported_models_base.ClipTarget(comfy.text_encoders.z_image.ZImageTokenizer, comfy.text_encoders.z_image.te(**hunyuan_detect))
+
 class WAN21_T2V(supported_models_base.BASE):
     unet_config = {
         "image_model": "wan2.1",
@@ -1374,6 +1424,55 @@ class HunyuanImage21Refiner(HunyuanVideo):
         out = model_base.HunyuanImage21Refiner(self, device=device)
         return out
 
-models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage]
+class HunyuanVideo15(HunyuanVideo):
+    unet_config = {
+        "image_model": "hunyuan_video",
+        "vision_in_dim": 1152,
+    }
+
+    sampling_settings = {
+        "shift": 7.0,
+    }
+    memory_usage_factor = 4.0 #TODO
+    supported_inference_dtypes = [torch.float16, torch.bfloat16, torch.float32]
+
+    latent_format = latent_formats.HunyuanVideo15
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.HunyuanVideo15(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        pref = self.text_encoder_key_prefix[0]
+        hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_7b.transformer.".format(pref))
+        return supported_models_base.ClipTarget(comfy.text_encoders.hunyuan_video.HunyuanVideo15Tokenizer, comfy.text_encoders.hunyuan_image.te(**hunyuan_detect))
+
+
+class HunyuanVideo15_SR_Distilled(HunyuanVideo):
+    unet_config = {
+        "image_model": "hunyuan_video",
+        "vision_in_dim": 1152,
+        "in_channels": 98,
+    }
+
+    sampling_settings = {
+        "shift": 2.0,
+    }
+    memory_usage_factor = 4.0 #TODO
+    supported_inference_dtypes = [torch.float16, torch.bfloat16, torch.float32]
+
+    latent_format = latent_formats.HunyuanVideo15
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.HunyuanVideo15_SR_Distilled(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        pref = self.text_encoder_key_prefix[0]
+        hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_7b.transformer.".format(pref))
+        return supported_models_base.ClipTarget(comfy.text_encoders.hunyuan_video.HunyuanVideo15Tokenizer, comfy.text_encoders.hunyuan_image.te(**hunyuan_detect))
+
+models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Flux2]
+
 
 models += [SVD_img2vid]

comfy/text_encoders/flux.py

@@ -1,10 +1,13 @@
 from comfy import sd1_clip
 import comfy.text_encoders.t5
 import comfy.text_encoders.sd3_clip
+import comfy.text_encoders.llama
 import comfy.model_management
-from transformers import T5TokenizerFast
+from transformers import T5TokenizerFast, LlamaTokenizerFast
 import torch
 import os
+import json
+import base64
 
 class T5XXLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -68,3 +71,106 @@ def flux_clip(dtype_t5=None, t5xxl_scaled_fp8=None):
                 model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8
             super().__init__(dtype_t5=dtype_t5, device=device, dtype=dtype, model_options=model_options)
     return FluxClipModel_
+
+def load_mistral_tokenizer(data):
+    if torch.is_tensor(data):
+        data = data.numpy().tobytes()
+
+    try:
+        from transformers.integrations.mistral import MistralConverter
+    except ModuleNotFoundError:
+        from transformers.models.pixtral.convert_pixtral_weights_to_hf import MistralConverter
+
+    mistral_vocab = json.loads(data)
+
+    special_tokens = {}
+    vocab = {}
+
+    max_vocab = mistral_vocab["config"]["default_vocab_size"]
+    max_vocab -= len(mistral_vocab["special_tokens"])
+
+    for w in mistral_vocab["vocab"]:
+        r = w["rank"]
+        if r >= max_vocab:
+            continue
+
+        vocab[base64.b64decode(w["token_bytes"])] = r
+
+    for w in mistral_vocab["special_tokens"]:
+        if "token_bytes" in w:
+            special_tokens[base64.b64decode(w["token_bytes"])] = w["rank"]
+        else:
+            special_tokens[w["token_str"]] = w["rank"]
+
+    all_special = []
+    for v in special_tokens:
+        all_special.append(v)
+
+    special_tokens.update(vocab)
+    vocab = special_tokens
+    return {"tokenizer_object": MistralConverter(vocab=vocab, additional_special_tokens=all_special).converted(), "legacy": False}
+
+class MistralTokenizerClass:
+    @staticmethod
+    def from_pretrained(path, **kwargs):
+        return LlamaTokenizerFast(**kwargs)
+
+class Mistral3Tokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        self.tekken_data = tokenizer_data.get("tekken_model", None)
+        super().__init__("", pad_with_end=False, embedding_size=5120, embedding_key='mistral3_24b', tokenizer_class=MistralTokenizerClass, has_end_token=False, pad_to_max_length=False, pad_token=11, max_length=99999999, min_length=1, pad_left=True, tokenizer_args=load_mistral_tokenizer(self.tekken_data), tokenizer_data=tokenizer_data)
+
+    def state_dict(self):
+        return {"tekken_model": self.tekken_data}
+
+class Flux2Tokenizer(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="mistral3_24b", tokenizer=Mistral3Tokenizer)
+        self.llama_template = '[SYSTEM_PROMPT]You are an AI that reasons about image descriptions. You give structured responses focusing on object relationships, object\nattribution and actions without speculation.[/SYSTEM_PROMPT][INST]{}[/INST]'
+
+    def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, **kwargs):
+        if llama_template is None:
+            llama_text = self.llama_template.format(text)
+        else:
+            llama_text = llama_template.format(text)
+
+        tokens = super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, disable_weights=True, **kwargs)
+        return tokens
+
+class Mistral3_24BModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer=[10, 20, 30], layer_idx=None, dtype=None, attention_mask=True, model_options={}):
+        textmodel_json_config = {}
+        num_layers = model_options.get("num_layers", None)
+        if num_layers is not None:
+            textmodel_json_config["num_hidden_layers"] = num_layers
+            if num_layers < 40:
+                textmodel_json_config["final_norm"] = False
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 1, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Mistral3Small24B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
+class Flux2TEModel(sd1_clip.SD1ClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}, name="mistral3_24b", clip_model=Mistral3_24BModel):
+        super().__init__(device=device, dtype=dtype, name=name, clip_model=clip_model, model_options=model_options)
+
+    def encode_token_weights(self, token_weight_pairs):
+        out, pooled, extra = super().encode_token_weights(token_weight_pairs)
+
+        out = torch.stack((out[:, 0], out[:, 1], out[:, 2]), dim=1)
+        out = out.movedim(1, 2)
+        out = out.reshape(out.shape[0], out.shape[1], -1)
+        return out, pooled, extra
+
+def flux2_te(dtype_llama=None, llama_scaled_fp8=None, llama_quantization_metadata=None, pruned=False):
+    class Flux2TEModel_(Flux2TEModel):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options:
+                model_options = model_options.copy()
+                model_options["scaled_fp8"] = llama_scaled_fp8
+            if dtype_llama is not None:
+                dtype = dtype_llama
+            if llama_quantization_metadata is not None:
+                model_options["quantization_metadata"] = llama_quantization_metadata
+            if pruned:
+                model_options = model_options.copy()
+                model_options["num_layers"] = 30
+            super().__init__(device=device, dtype=dtype, model_options=model_options)
+    return Flux2TEModel_

comfy/text_encoders/hunyuan_video.py

@@ -1,6 +1,7 @@
 from comfy import sd1_clip
 import comfy.model_management
 import comfy.text_encoders.llama
+from .hunyuan_image import HunyuanImageTokenizer
 from transformers import LlamaTokenizerFast
 import torch
 import os
@@ -17,6 +18,9 @@ def llama_detect(state_dict, prefix=""):
     if scaled_fp8_key in state_dict:
         out["llama_scaled_fp8"] = state_dict[scaled_fp8_key].dtype
 
+    if "_quantization_metadata" in state_dict:
+        out["llama_quantization_metadata"] = state_dict["_quantization_metadata"]
+
     return out
 
 
@@ -73,6 +77,14 @@ class HunyuanVideoTokenizer:
         return {}
 
 
+class HunyuanVideo15Tokenizer(HunyuanImageTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
+        self.llama_template = "<|im_start|>system\nYou are a helpful assistant. Describe the video by detailing the following aspects:\n1. The main content and theme of the video.\n2. The color, shape, size, texture, quantity, text, and spatial relationships of the objects.\n3. Actions, events, behaviors temporal relationships, physical movement changes of the objects.\n4. background environment, light, style and atmosphere.\n5. camera angles, movements, and transitions used in the video.<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n"
+
+    def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
+        return super().tokenize_with_weights(text, return_word_ids, prevent_empty_text=True, **kwargs)
+
 class HunyuanVideoClipModel(torch.nn.Module):
     def __init__(self, dtype_llama=None, device="cpu", dtype=None, model_options={}):
         super().__init__()

comfy/text_encoders/llama.py

@@ -34,6 +34,28 @@ class Llama2Config:
     rope_scale = None
     final_norm: bool = True
 
+@dataclass
+class Mistral3Small24BConfig:
+    vocab_size: int = 131072
+    hidden_size: int = 5120
+    intermediate_size: int = 32768
+    num_hidden_layers: int = 40
+    num_attention_heads: int = 32
+    num_key_value_heads: int = 8
+    max_position_embeddings: int = 8192
+    rms_norm_eps: float = 1e-5
+    rope_theta: float = 1000000000.0
+    transformer_type: str = "llama"
+    head_dim = 128
+    rms_norm_add = False
+    mlp_activation = "silu"
+    qkv_bias = False
+    rope_dims = None
+    q_norm = None
+    k_norm = None
+    rope_scale = None
+    final_norm: bool = True
+
 @dataclass
 class Qwen25_3BConfig:
     vocab_size: int = 151936
@@ -56,6 +78,28 @@ class Qwen25_3BConfig:
     rope_scale = None
     final_norm: bool = True
 
+@dataclass
+class Qwen3_4BConfig:
+    vocab_size: int = 151936
+    hidden_size: int = 2560
+    intermediate_size: int = 9728
+    num_hidden_layers: int = 36
+    num_attention_heads: int = 32
+    num_key_value_heads: int = 8
+    max_position_embeddings: int = 40960
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 1000000.0
+    transformer_type: str = "llama"
+    head_dim = 128
+    rms_norm_add = False
+    mlp_activation = "silu"
+    qkv_bias = False
+    rope_dims = None
+    q_norm = "gemma3"
+    k_norm = "gemma3"
+    rope_scale = None
+    final_norm: bool = True
+
 @dataclass
 class Qwen25_7BVLI_Config:
     vocab_size: int = 152064
@@ -412,8 +456,12 @@ class Llama2_(nn.Module):
 
         intermediate = None
         all_intermediate = None
+        only_layers = None
         if intermediate_output is not None:
-            if intermediate_output == "all":
+            if isinstance(intermediate_output, list):
+                all_intermediate = []
+                only_layers = set(intermediate_output)
+            elif intermediate_output == "all":
                 all_intermediate = []
                 intermediate_output = None
             elif intermediate_output < 0:
@@ -421,7 +469,8 @@ class Llama2_(nn.Module):
 
         for i, layer in enumerate(self.layers):
             if all_intermediate is not None:
-                all_intermediate.append(x.unsqueeze(1).clone())
+                if only_layers is None or (i in only_layers):
+                    all_intermediate.append(x.unsqueeze(1).clone())
             x = layer(
                 x=x,
                 attention_mask=mask,
@@ -435,7 +484,8 @@ class Llama2_(nn.Module):
             x = self.norm(x)
 
         if all_intermediate is not None:
-            all_intermediate.append(x.unsqueeze(1).clone())
+            if only_layers is None or ((i + 1) in only_layers):
+                all_intermediate.append(x.unsqueeze(1).clone())
 
         if all_intermediate is not None:
             intermediate = torch.cat(all_intermediate, dim=1)
@@ -465,6 +515,15 @@ class Llama2(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
+class Mistral3Small24B(BaseLlama, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Mistral3Small24BConfig(**config_dict)
+        self.num_layers = config.num_hidden_layers
+
+        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.dtype = dtype
+
 class Qwen25_3B(BaseLlama, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
@@ -474,6 +533,15 @@ class Qwen25_3B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
+class Qwen3_4B(BaseLlama, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Qwen3_4BConfig(**config_dict)
+        self.num_layers = config.num_hidden_layers
+
+        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.dtype = dtype
+
 class Qwen25_7BVLI(BaseLlama, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()

comfy/text_encoders/qwen_image.py

@@ -17,12 +17,14 @@ class QwenImageTokenizer(sd1_clip.SD1Tokenizer):
         self.llama_template = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n"
         self.llama_template_images = "<|im_start|>system\nDescribe the key features of the input image (color, shape, size, texture, objects, background), then explain how the user's text instruction should alter or modify the image. Generate a new image that meets the user's requirements while maintaining consistency with the original input where appropriate.<|im_end|>\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>{}<|im_end|>\n<|im_start|>assistant\n"
 
-    def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, images=[], **kwargs):
+    def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, images=[], prevent_empty_text=False, **kwargs):
         skip_template = False
         if text.startswith('<|im_start|>'):
             skip_template = True
         if text.startswith('<|start_header_id|>'):
             skip_template = True
+        if prevent_empty_text and text == '':
+            text = ' '
 
         if skip_template:
             llama_text = text

comfy/text_encoders/z_image.py

@@ -0,0 +1,48 @@
+from transformers import Qwen2Tokenizer
+import comfy.text_encoders.llama
+from comfy import sd1_clip
+import os
+
+class Qwen3Tokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data)
+
+
+class ZImageTokenizer(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="qwen3_4b", tokenizer=Qwen3Tokenizer)
+        self.llama_template = "<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n"
+
+    def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, **kwargs):
+        if llama_template is None:
+            llama_text = self.llama_template.format(text)
+        else:
+            llama_text = llama_template.format(text)
+
+        tokens = super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, disable_weights=True, **kwargs)
+        return tokens
+
+
+class Qwen3_4BModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="hidden", layer_idx=-2, dtype=None, attention_mask=True, model_options={}):
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_4B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
+
+class ZImageTEModel(sd1_clip.SD1ClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, dtype=dtype, name="qwen3_4b", clip_model=Qwen3_4BModel, model_options=model_options)
+
+
+def te(dtype_llama=None, llama_scaled_fp8=None, llama_quantization_metadata=None):
+    class ZImageTEModel_(ZImageTEModel):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options:
+                model_options = model_options.copy()
+                model_options["scaled_fp8"] = llama_scaled_fp8
+            if dtype_llama is not None:
+                dtype = dtype_llama
+            if llama_quantization_metadata is not None:
+                model_options["quantization_metadata"] = llama_quantization_metadata
+            super().__init__(device=device, dtype=dtype, model_options=model_options)
+    return ZImageTEModel_

comfy/weight_adapter/lora.py

@@ -194,6 +194,7 @@ class LoRAAdapter(WeightAdapterBase):
             lora_diff = torch.mm(
                 mat1.flatten(start_dim=1), mat2.flatten(start_dim=1)
             ).reshape(weight.shape)
+            del mat1, mat2
             if dora_scale is not None:
                 weight = weight_decompose(
                     dora_scale,

comfy_api/internal/async_to_sync.py

@@ -8,7 +8,7 @@ import os
 import textwrap
 import threading
 from enum import Enum
-from typing import Optional, Type, get_origin, get_args
+from typing import Optional, Type, get_origin, get_args, get_type_hints
 
 
 class TypeTracker:
@@ -220,11 +220,18 @@ class AsyncToSyncConverter:
             self._async_instance = async_class(*args, **kwargs)
 
             # Handle annotated class attributes (like execution: Execution)
-            # Get all annotations from the class hierarchy
-            all_annotations = {}
-            for base_class in reversed(inspect.getmro(async_class)):
-                if hasattr(base_class, "__annotations__"):
-                    all_annotations.update(base_class.__annotations__)
+            # Get all annotations from the class hierarchy and resolve string annotations
+            try:
+                # get_type_hints resolves string annotations to actual type objects
+                # This handles classes using 'from __future__ import annotations'
+                all_annotations = get_type_hints(async_class)
+            except Exception:
+                # Fallback to raw annotations if get_type_hints fails
+                # (e.g., for undefined forward references)
+                all_annotations = {}
+                for base_class in reversed(inspect.getmro(async_class)):
+                    if hasattr(base_class, "__annotations__"):
+                        all_annotations.update(base_class.__annotations__)
 
             # For each annotated attribute, check if it needs to be created or wrapped
             for attr_name, attr_type in all_annotations.items():
@@ -625,15 +632,19 @@ class AsyncToSyncConverter:
         """Extract class attributes that are classes themselves."""
         class_attributes = []
 
+        # Get resolved type hints to handle string annotations
+        try:
+            type_hints = get_type_hints(async_class)
+        except Exception:
+            type_hints = {}
+
         # Look for class attributes that are classes
         for name, attr in sorted(inspect.getmembers(async_class)):
             if isinstance(attr, type) and not name.startswith("_"):
                 class_attributes.append((name, attr))
-            elif (
-                hasattr(async_class, "__annotations__")
-                and name in async_class.__annotations__
-            ):
-                annotation = async_class.__annotations__[name]
+            elif name in type_hints:
+                # Use resolved type hint instead of raw annotation
+                annotation = type_hints[name]
                 if isinstance(annotation, type):
                     class_attributes.append((name, annotation))
 
@@ -908,11 +919,15 @@ class AsyncToSyncConverter:
             attribute_mappings = {}
 
             # First check annotations for typed attributes (including from parent classes)
-            # Collect all annotations from the class hierarchy
-            all_annotations = {}
-            for base_class in reversed(inspect.getmro(async_class)):
-                if hasattr(base_class, "__annotations__"):
-                    all_annotations.update(base_class.__annotations__)
+            # Resolve string annotations to actual types
+            try:
+                all_annotations = get_type_hints(async_class)
+            except Exception:
+                # Fallback to raw annotations
+                all_annotations = {}
+                for base_class in reversed(inspect.getmro(async_class)):
+                    if hasattr(base_class, "__annotations__"):
+                        all_annotations.update(base_class.__annotations__)
 
             for attr_name, attr_type in sorted(all_annotations.items()):
                 for class_name, class_type in class_attributes:

comfy_api/latest/_input/video_types.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 from abc import ABC, abstractmethod
+from fractions import Fraction
 from typing import Optional, Union, IO
 import io
 import av
@@ -72,6 +73,33 @@ class VideoInput(ABC):
         frame_count = components.images.shape[0]
         return float(frame_count / components.frame_rate)
 
+    def get_frame_count(self) -> int:
+        """
+        Returns the number of frames in the video.
+
+        Default implementation uses :meth:`get_components`, which may require
+        loading all frames into memory. File-based implementations should
+        override this method and use container/stream metadata instead.
+
+        Returns:
+            Total number of frames as an integer.
+        """
+        return int(self.get_components().images.shape[0])
+
+    def get_frame_rate(self) -> Fraction:
+        """
+        Returns the frame rate of the video.
+
+        Default implementation materializes the video into memory via
+        `get_components()`. Subclasses that can inspect the underlying
+        container (e.g. `VideoFromFile`) should override this with a more
+        efficient implementation.
+
+        Returns:
+            Frame rate as a Fraction.
+        """
+        return self.get_components().frame_rate
+
     def get_container_format(self) -> str:
         """
         Returns the container format of the video (e.g., 'mp4', 'mov', 'avi').

comfy_api/latest/_input_impl/video_types.py

@@ -121,6 +121,71 @@ class VideoFromFile(VideoInput):
 
         raise ValueError(f"Could not determine duration for file '{self.__file}'")
 
+    def get_frame_count(self) -> int:
+        """
+        Returns the number of frames in the video without materializing them as
+        torch tensors.
+        """
+        if isinstance(self.__file, io.BytesIO):
+            self.__file.seek(0)
+
+        with av.open(self.__file, mode="r") as container:
+            video_stream = self._get_first_video_stream(container)
+            # 1. Prefer the frames field if available
+            if video_stream.frames and video_stream.frames > 0:
+                return int(video_stream.frames)
+
+            # 2. Try to estimate from duration and average_rate using only metadata
+            if container.duration is not None and video_stream.average_rate:
+                duration_seconds = float(container.duration / av.time_base)
+                estimated_frames = int(round(duration_seconds * float(video_stream.average_rate)))
+                if estimated_frames > 0:
+                    return estimated_frames
+
+            if (
+                getattr(video_stream, "duration", None) is not None
+                and getattr(video_stream, "time_base", None) is not None
+                and video_stream.average_rate
+            ):
+                duration_seconds = float(video_stream.duration * video_stream.time_base)
+                estimated_frames = int(round(duration_seconds * float(video_stream.average_rate)))
+                if estimated_frames > 0:
+                    return estimated_frames
+
+            # 3. Last resort: decode frames and count them (streaming)
+            frame_count = 0
+            container.seek(0)
+            for packet in container.demux(video_stream):
+                for _ in packet.decode():
+                    frame_count += 1
+
+            if frame_count == 0:
+                raise ValueError(f"Could not determine frame count for file '{self.__file}'")
+            return frame_count
+
+    def get_frame_rate(self) -> Fraction:
+        """
+        Returns the average frame rate of the video using container metadata
+        without decoding all frames.
+        """
+        if isinstance(self.__file, io.BytesIO):
+            self.__file.seek(0)
+
+        with av.open(self.__file, mode="r") as container:
+            video_stream = self._get_first_video_stream(container)
+            # Preferred: use PyAV's average_rate (usually already a Fraction-like)
+            if video_stream.average_rate:
+                return Fraction(video_stream.average_rate)
+
+            # Fallback: estimate from frames + duration if available
+            if video_stream.frames and container.duration:
+                duration_seconds = float(container.duration / av.time_base)
+                if duration_seconds > 0:
+                    return Fraction(video_stream.frames / duration_seconds).limit_denominator()
+
+            # Last resort: match get_components_internal default
+            return Fraction(1)
+
     def get_container_format(self) -> str:
         """
         Returns the container format of the video (e.g., 'mp4', 'mov', 'avi').
@@ -238,6 +303,13 @@ class VideoFromFile(VideoInput):
                         packet.stream = stream_map[packet.stream]
                         output_container.mux(packet)
 
+    def _get_first_video_stream(self, container: InputContainer):
+        video_stream = next((s for s in container.streams if s.type == "video"), None)
+        if video_stream is None:
+            raise ValueError(f"No video stream found in file '{self.__file}'")
+        return video_stream
+
+
 class VideoFromComponents(VideoInput):
     """
     Class representing video input from tensors.

comfy_api/latest/_io.py

@@ -629,6 +629,10 @@ class UpscaleModel(ComfyTypeIO):
     if TYPE_CHECKING:
         Type = ImageModelDescriptor
 
+@comfytype(io_type="LATENT_UPSCALE_MODEL")
+class LatentUpscaleModel(ComfyTypeIO):
+    Type = Any
+
 @comfytype(io_type="AUDIO")
 class Audio(ComfyTypeIO):
     class AudioDict(TypedDict):

comfy_api_nodes/apis/bfl_api.py

@@ -70,6 +70,29 @@ class BFLFluxProGenerateRequest(BaseModel):
     # )
 
 
+class Flux2ProGenerateRequest(BaseModel):
+    prompt: str = Field(...)
+    width: int = Field(1024, description="Must be a multiple of 32.")
+    height: int = Field(768, description="Must be a multiple of 32.")
+    seed: int | None = Field(None)
+    prompt_upsampling: bool | None = Field(None)
+    input_image: str | None = Field(None, description="Base64 encoded image for image-to-image generation")
+    input_image_2: str | None = Field(None, description="Base64 encoded image for image-to-image generation")
+    input_image_3: str | None = Field(None, description="Base64 encoded image for image-to-image generation")
+    input_image_4: str | None = Field(None, description="Base64 encoded image for image-to-image generation")
+    input_image_5: str | None = Field(None, description="Base64 encoded image for image-to-image generation")
+    input_image_6: str | None = Field(None, description="Base64 encoded image for image-to-image generation")
+    input_image_7: str | None = Field(None, description="Base64 encoded image for image-to-image generation")
+    input_image_8: str | None = Field(None, description="Base64 encoded image for image-to-image generation")
+    input_image_9: str | None = Field(None, description="Base64 encoded image for image-to-image generation")
+    safety_tolerance: int | None = Field(
+        5, description="Tolerance level for input and output moderation. Value 0 being most strict.", ge=0, le=5
+    )
+    output_format: str | None = Field(
+        "png", description="Output format for the generated image. Can be 'jpeg' or 'png'."
+    )
+
+
 class BFLFluxKontextProGenerateRequest(BaseModel):
     prompt: str = Field(..., description='The text prompt for what you wannt to edit.')
     input_image: Optional[str] = Field(None, description='Image to edit in base64 format')
@@ -109,8 +132,9 @@ class BFLFluxProUltraGenerateRequest(BaseModel):
 
 
 class BFLFluxProGenerateResponse(BaseModel):
-    id: str = Field(..., description='The unique identifier for the generation task.')
-    polling_url: str = Field(..., description='URL to poll for the generation result.')
+    id: str = Field(..., description="The unique identifier for the generation task.")
+    polling_url: str = Field(..., description="URL to poll for the generation result.")
+    cost: float | None = Field(None, description="Price in cents")
 
 
 class BFLStatus(str, Enum):

comfy_api_nodes/apis/gemini_api.py

@@ -113,9 +113,9 @@ class GeminiGenerationConfig(BaseModel):
     maxOutputTokens: int | None = Field(None, ge=16, le=8192)
     seed: int | None = Field(None)
     stopSequences: list[str] | None = Field(None)
-    temperature: float | None = Field(1, ge=0.0, le=2.0)
-    topK: int | None = Field(40, ge=1)
-    topP: float | None = Field(0.95, ge=0.0, le=1.0)
+    temperature: float | None = Field(None, ge=0.0, le=2.0)
+    topK: int | None = Field(None, ge=1)
+    topP: float | None = Field(None, ge=0.0, le=1.0)
 
 
 class GeminiImageConfig(BaseModel):

comfy_api_nodes/nodes_bfl.py

@@ -1,7 +1,7 @@
 from inspect import cleandoc
-from typing import Optional
 
 import torch
+from pydantic import BaseModel
 from typing_extensions import override
 
 from comfy_api.latest import IO, ComfyExtension
@@ -9,15 +9,16 @@ from comfy_api_nodes.apis.bfl_api import (
     BFLFluxExpandImageRequest,
     BFLFluxFillImageRequest,
     BFLFluxKontextProGenerateRequest,
-    BFLFluxProGenerateRequest,
     BFLFluxProGenerateResponse,
     BFLFluxProUltraGenerateRequest,
     BFLFluxStatusResponse,
     BFLStatus,
+    Flux2ProGenerateRequest,
 )
 from comfy_api_nodes.util import (
     ApiEndpoint,
     download_url_to_image_tensor,
+    get_number_of_images,
     poll_op,
     resize_mask_to_image,
     sync_op,
@@ -116,7 +117,7 @@ class FluxProUltraImageNode(IO.ComfyNode):
         prompt_upsampling: bool = False,
         raw: bool = False,
         seed: int = 0,
-        image_prompt: Optional[torch.Tensor] = None,
+        image_prompt: torch.Tensor | None = None,
         image_prompt_strength: float = 0.1,
     ) -> IO.NodeOutput:
         if image_prompt is None:
@@ -230,7 +231,7 @@ class FluxKontextProImageNode(IO.ComfyNode):
         aspect_ratio: str,
         guidance: float,
         steps: int,
-        input_image: Optional[torch.Tensor] = None,
+        input_image: torch.Tensor | None = None,
         seed=0,
         prompt_upsampling=False,
     ) -> IO.NodeOutput:
@@ -280,124 +281,6 @@ class FluxKontextMaxImageNode(FluxKontextProImageNode):
     DISPLAY_NAME = "Flux.1 Kontext [max] Image"
 
 
-class FluxProImageNode(IO.ComfyNode):
-    """
-    Generates images synchronously based on prompt and resolution.
-    """
-
-    @classmethod
-    def define_schema(cls) -> IO.Schema:
-        return IO.Schema(
-            node_id="FluxProImageNode",
-            display_name="Flux 1.1 [pro] Image",
-            category="api node/image/BFL",
-            description=cleandoc(cls.__doc__ or ""),
-            inputs=[
-                IO.String.Input(
-                    "prompt",
-                    multiline=True,
-                    default="",
-                    tooltip="Prompt for the image generation",
-                ),
-                IO.Boolean.Input(
-                    "prompt_upsampling",
-                    default=False,
-                    tooltip="Whether to perform upsampling on the prompt. "
-                    "If active, automatically modifies the prompt for more creative generation, "
-                    "but results are nondeterministic (same seed will not produce exactly the same result).",
-                ),
-                IO.Int.Input(
-                    "width",
-                    default=1024,
-                    min=256,
-                    max=1440,
-                    step=32,
-                ),
-                IO.Int.Input(
-                    "height",
-                    default=768,
-                    min=256,
-                    max=1440,
-                    step=32,
-                ),
-                IO.Int.Input(
-                    "seed",
-                    default=0,
-                    min=0,
-                    max=0xFFFFFFFFFFFFFFFF,
-                    control_after_generate=True,
-                    tooltip="The random seed used for creating the noise.",
-                ),
-                IO.Image.Input(
-                    "image_prompt",
-                    optional=True,
-                ),
-                # "image_prompt_strength": (
-                #     IO.FLOAT,
-                #     {
-                #         "default": 0.1,
-                #         "min": 0.0,
-                #         "max": 1.0,
-                #         "step": 0.01,
-                #         "tooltip": "Blend between the prompt and the image prompt.",
-                #     },
-                # ),
-            ],
-            outputs=[IO.Image.Output()],
-            hidden=[
-                IO.Hidden.auth_token_comfy_org,
-                IO.Hidden.api_key_comfy_org,
-                IO.Hidden.unique_id,
-            ],
-            is_api_node=True,
-        )
-
-    @classmethod
-    async def execute(
-        cls,
-        prompt: str,
-        prompt_upsampling,
-        width: int,
-        height: int,
-        seed=0,
-        image_prompt=None,
-        # image_prompt_strength=0.1,
-    ) -> IO.NodeOutput:
-        image_prompt = image_prompt if image_prompt is None else tensor_to_base64_string(image_prompt)
-        initial_response = await sync_op(
-            cls,
-            ApiEndpoint(
-                path="/proxy/bfl/flux-pro-1.1/generate",
-                method="POST",
-            ),
-            response_model=BFLFluxProGenerateResponse,
-            data=BFLFluxProGenerateRequest(
-                prompt=prompt,
-                prompt_upsampling=prompt_upsampling,
-                width=width,
-                height=height,
-                seed=seed,
-                image_prompt=image_prompt,
-            ),
-        )
-        response = await poll_op(
-            cls,
-            ApiEndpoint(initial_response.polling_url),
-            response_model=BFLFluxStatusResponse,
-            status_extractor=lambda r: r.status,
-            progress_extractor=lambda r: r.progress,
-            completed_statuses=[BFLStatus.ready],
-            failed_statuses=[
-                BFLStatus.request_moderated,
-                BFLStatus.content_moderated,
-                BFLStatus.error,
-                BFLStatus.task_not_found,
-            ],
-            queued_statuses=[],
-        )
-        return IO.NodeOutput(await download_url_to_image_tensor(response.result["sample"]))
-
-
 class FluxProExpandNode(IO.ComfyNode):
     """
     Outpaints image based on prompt.
@@ -640,16 +523,125 @@ class FluxProFillNode(IO.ComfyNode):
         return IO.NodeOutput(await download_url_to_image_tensor(response.result["sample"]))
 
 
+class Flux2ProImageNode(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="Flux2ProImageNode",
+            display_name="Flux.2 [pro] Image",
+            category="api node/image/BFL",
+            description="Generates images synchronously based on prompt and resolution.",
+            inputs=[
+                IO.String.Input(
+                    "prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Prompt for the image generation or edit",
+                ),
+                IO.Int.Input(
+                    "width",
+                    default=1024,
+                    min=256,
+                    max=2048,
+                    step=32,
+                ),
+                IO.Int.Input(
+                    "height",
+                    default=768,
+                    min=256,
+                    max=2048,
+                    step=32,
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=0xFFFFFFFFFFFFFFFF,
+                    control_after_generate=True,
+                    tooltip="The random seed used for creating the noise.",
+                ),
+                IO.Boolean.Input(
+                    "prompt_upsampling",
+                    default=False,
+                    tooltip="Whether to perform upsampling on the prompt. "
+                    "If active, automatically modifies the prompt for more creative generation, "
+                    "but results are nondeterministic (same seed will not produce exactly the same result).",
+                ),
+                IO.Image.Input("images", optional=True, tooltip="Up to 4 images to be used as references."),
+            ],
+            outputs=[IO.Image.Output()],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        prompt: str,
+        width: int,
+        height: int,
+        seed: int,
+        prompt_upsampling: bool,
+        images: torch.Tensor | None = None,
+    ) -> IO.NodeOutput:
+        reference_images = {}
+        if images is not None:
+            if get_number_of_images(images) > 9:
+                raise ValueError("The current maximum number of supported images is 9.")
+            for image_index in range(images.shape[0]):
+                key_name = f"input_image_{image_index + 1}" if image_index else "input_image"
+                reference_images[key_name] = tensor_to_base64_string(images[image_index], total_pixels=2048 * 2048)
+        initial_response = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/bfl/flux-2-pro/generate", method="POST"),
+            response_model=BFLFluxProGenerateResponse,
+            data=Flux2ProGenerateRequest(
+                prompt=prompt,
+                width=width,
+                height=height,
+                seed=seed,
+                prompt_upsampling=prompt_upsampling,
+                **reference_images,
+            ),
+        )
+
+        def price_extractor(_r: BaseModel) -> float | None:
+            return None if initial_response.cost is None else initial_response.cost / 100
+
+        response = await poll_op(
+            cls,
+            ApiEndpoint(initial_response.polling_url),
+            response_model=BFLFluxStatusResponse,
+            status_extractor=lambda r: r.status,
+            progress_extractor=lambda r: r.progress,
+            price_extractor=price_extractor,
+            completed_statuses=[BFLStatus.ready],
+            failed_statuses=[
+                BFLStatus.request_moderated,
+                BFLStatus.content_moderated,
+                BFLStatus.error,
+                BFLStatus.task_not_found,
+            ],
+            queued_statuses=[],
+        )
+        return IO.NodeOutput(await download_url_to_image_tensor(response.result["sample"]))
+
+
 class BFLExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[IO.ComfyNode]]:
         return [
             FluxProUltraImageNode,
-            # FluxProImageNode,
             FluxKontextProImageNode,
             FluxKontextMaxImageNode,
             FluxProExpandNode,
             FluxProFillNode,
+            Flux2ProImageNode,
         ]
 
 

comfy_api_nodes/nodes_gemini.py

@@ -104,14 +104,14 @@ def get_parts_by_type(response: GeminiGenerateContentResponse, part_type: Litera
         List of response parts matching the requested type.
     """
     if response.candidates is None:
-        if response.promptFeedback.blockReason:
+        if response.promptFeedback and response.promptFeedback.blockReason:
             feedback = response.promptFeedback
             raise ValueError(
                 f"Gemini API blocked the request. Reason: {feedback.blockReason} ({feedback.blockReasonMessage})"
             )
-        raise NotImplementedError(
-            "Gemini returned no response candidates. "
-            "Please report to ComfyUI repository with the example of workflow to reproduce this."
+        raise ValueError(
+            "Gemini API returned no response candidates. If you are using the `IMAGE` modality, "
+            "try changing it to `IMAGE+TEXT` to view the model's reasoning and understand why image generation failed."
         )
     parts = []
     for part in response.candidates[0].content.parts:
@@ -182,11 +182,12 @@ def calculate_tokens_price(response: GeminiGenerateContentResponse) -> float | N
     else:
         return None
     final_price = response.usageMetadata.promptTokenCount * input_tokens_price
-    for i in response.usageMetadata.candidatesTokensDetails:
-        if i.modality == Modality.IMAGE:
-            final_price += output_image_tokens_price * i.tokenCount  # for Nano Banana models
-        else:
-            final_price += output_text_tokens_price * i.tokenCount
+    if response.usageMetadata.candidatesTokensDetails:
+        for i in response.usageMetadata.candidatesTokensDetails:
+            if i.modality == Modality.IMAGE:
+                final_price += output_image_tokens_price * i.tokenCount  # for Nano Banana models
+            else:
+                final_price += output_text_tokens_price * i.tokenCount
     if response.usageMetadata.thoughtsTokenCount:
         final_price += output_text_tokens_price * response.usageMetadata.thoughtsTokenCount
     return final_price / 1_000_000.0
@@ -645,7 +646,7 @@ class GeminiImage2(IO.ComfyNode):
                     options=["auto", "1:1", "2:3", "3:2", "3:4", "4:3", "4:5", "5:4", "9:16", "16:9", "21:9"],
                     default="auto",
                     tooltip="If set to 'auto', matches your input image's aspect ratio; "
-                    "if no image is provided, generates a 1:1 square.",
+                    "if no image is provided, a 16:9 square is usually generated.",
                 ),
                 IO.Combo.Input(
                     "resolution",

comfy_api_nodes/nodes_topaz.py

@@ -5,8 +5,7 @@ import aiohttp
 import torch
 from typing_extensions import override
 
-from comfy_api.input.video_types import VideoInput
-from comfy_api.latest import IO, ComfyExtension
+from comfy_api.latest import IO, ComfyExtension, Input
 from comfy_api_nodes.apis import topaz_api
 from comfy_api_nodes.util import (
     ApiEndpoint,
@@ -282,7 +281,7 @@ class TopazVideoEnhance(IO.ComfyNode):
     @classmethod
     async def execute(
         cls,
-        video: VideoInput,
+        video: Input.Video,
         upscaler_enabled: bool,
         upscaler_model: str,
         upscaler_resolution: str,
@@ -297,12 +296,10 @@ class TopazVideoEnhance(IO.ComfyNode):
     ) -> IO.NodeOutput:
         if upscaler_enabled is False and interpolation_enabled is False:
             raise ValueError("There is nothing to do: both upscaling and interpolation are disabled.")
-        src_width, src_height = video.get_dimensions()
-        video_components = video.get_components()
-        src_frame_rate = int(video_components.frame_rate)
-        duration_sec = video.get_duration()
-        estimated_frames = int(duration_sec * src_frame_rate)
         validate_container_format_is_mp4(video)
+        src_width, src_height = video.get_dimensions()
+        src_frame_rate = int(video.get_frame_rate())
+        duration_sec = video.get_duration()
         src_video_stream = video.get_stream_source()
         target_width = src_width
         target_height = src_height
@@ -338,7 +335,7 @@ class TopazVideoEnhance(IO.ComfyNode):
                     container="mp4",
                     size=get_fs_object_size(src_video_stream),
                     duration=int(duration_sec),
-                    frameCount=estimated_frames,
+                    frameCount=video.get_frame_count(),
                     frameRate=src_frame_rate,
                     resolution=topaz_api.Resolution(width=src_width, height=src_height),
                 ),

comfy_api_nodes/util/__init__.py

@@ -36,6 +36,7 @@ from .upload_helpers import (
     upload_video_to_comfyapi,
 )
 from .validation_utils import (
+    get_image_dimensions,
     get_number_of_images,
     validate_aspect_ratio_string,
     validate_audio_duration,
@@ -82,6 +83,7 @@ __all__ = [
     "trim_video",
     "video_to_base64_string",
     # Validation utilities
+    "get_image_dimensions",
     "get_number_of_images",
     "validate_aspect_ratio_string",
     "validate_audio_duration",

comfy_extras/nodes_flux.py

@@ -2,7 +2,10 @@ import node_helpers
 import comfy.utils
 from typing_extensions import override
 from comfy_api.latest import ComfyExtension, io
-
+import comfy.model_management
+import torch
+import math
+import nodes
 
 class CLIPTextEncodeFlux(io.ComfyNode):
     @classmethod
@@ -30,6 +33,27 @@ class CLIPTextEncodeFlux(io.ComfyNode):
 
     encode = execute  # TODO: remove
 
+class EmptyFlux2LatentImage(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="EmptyFlux2LatentImage",
+            display_name="Empty Flux 2 Latent",
+            category="latent",
+            inputs=[
+                io.Int.Input("width", default=1024, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("height", default=1024, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+            ],
+            outputs=[
+                io.Latent.Output(),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, width, height, batch_size=1) -> io.NodeOutput:
+        latent = torch.zeros([batch_size, 128, height // 16, width // 16], device=comfy.model_management.intermediate_device())
+        return io.NodeOutput({"samples": latent})
 
 class FluxGuidance(io.ComfyNode):
     @classmethod
@@ -154,6 +178,58 @@ class FluxKontextMultiReferenceLatentMethod(io.ComfyNode):
     append = execute  # TODO: remove
 
 
+def generalized_time_snr_shift(t, mu: float, sigma: float):
+    return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma)
+
+
+def compute_empirical_mu(image_seq_len: int, num_steps: int) -> float:
+    a1, b1 = 8.73809524e-05, 1.89833333
+    a2, b2 = 0.00016927, 0.45666666
+
+    if image_seq_len > 4300:
+        mu = a2 * image_seq_len + b2
+        return float(mu)
+
+    m_200 = a2 * image_seq_len + b2
+    m_10 = a1 * image_seq_len + b1
+
+    a = (m_200 - m_10) / 190.0
+    b = m_200 - 200.0 * a
+    mu = a * num_steps + b
+
+    return float(mu)
+
+
+def get_schedule(num_steps: int, image_seq_len: int) -> list[float]:
+    mu = compute_empirical_mu(image_seq_len, num_steps)
+    timesteps = torch.linspace(1, 0, num_steps + 1)
+    timesteps = generalized_time_snr_shift(timesteps, mu, 1.0)
+    return timesteps
+
+
+class Flux2Scheduler(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="Flux2Scheduler",
+            category="sampling/custom_sampling/schedulers",
+            inputs=[
+                io.Int.Input("steps", default=20, min=1, max=4096),
+                io.Int.Input("width", default=1024, min=16, max=nodes.MAX_RESOLUTION, step=1),
+                io.Int.Input("height", default=1024, min=16, max=nodes.MAX_RESOLUTION, step=1),
+            ],
+            outputs=[
+                io.Sigmas.Output(),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, steps, width, height) -> io.NodeOutput:
+        seq_len = (width * height / (16 * 16))
+        sigmas = get_schedule(steps, round(seq_len))
+        return io.NodeOutput(sigmas)
+
+
 class FluxExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[io.ComfyNode]]:
@@ -163,6 +239,8 @@ class FluxExtension(ComfyExtension):
             FluxDisableGuidance,
             FluxKontextImageScale,
             FluxKontextMultiReferenceLatentMethod,
+            EmptyFlux2LatentImage,
+            Flux2Scheduler,
         ]
 
 

comfy_extras/nodes_hunyuan.py

@@ -4,7 +4,8 @@ import torch
 import comfy.model_management
 from typing_extensions import override
 from comfy_api.latest import ComfyExtension, io
-
+from comfy.ldm.hunyuan_video.upsampler import HunyuanVideo15SRModel
+import folder_paths
 
 class CLIPTextEncodeHunyuanDiT(io.ComfyNode):
     @classmethod
@@ -37,6 +38,7 @@ class EmptyHunyuanLatentVideo(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="EmptyHunyuanLatentVideo",
+            display_name="Empty HunyuanVideo 1.0 Latent",
             category="latent/video",
             inputs=[
                 io.Int.Input("width", default=848, min=16, max=nodes.MAX_RESOLUTION, step=16),
@@ -57,6 +59,198 @@ class EmptyHunyuanLatentVideo(io.ComfyNode):
     generate = execute  # TODO: remove
 
 
+class EmptyHunyuanVideo15Latent(EmptyHunyuanLatentVideo):
+    @classmethod
+    def define_schema(cls):
+        schema = super().define_schema()
+        schema.node_id = "EmptyHunyuanVideo15Latent"
+        schema.display_name = "Empty HunyuanVideo 1.5 Latent"
+        return schema
+
+    @classmethod
+    def execute(cls, width, height, length, batch_size=1) -> io.NodeOutput:
+        # Using scale factor of 16 instead of 8
+        latent = torch.zeros([batch_size, 32, ((length - 1) // 4) + 1, height // 16, width // 16], device=comfy.model_management.intermediate_device())
+        return io.NodeOutput({"samples": latent})
+
+
+class HunyuanVideo15ImageToVideo(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="HunyuanVideo15ImageToVideo",
+            category="conditioning/video_models",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Int.Input("width", default=848, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("length", default=33, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.Image.Input("start_image", optional=True),
+                io.ClipVisionOutput.Input("clip_vision_output", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, positive, negative, vae, width, height, length, batch_size, start_image=None, clip_vision_output=None) -> io.NodeOutput:
+        latent = torch.zeros([batch_size, 32, ((length - 1) // 4) + 1, height // 16, width // 16], device=comfy.model_management.intermediate_device())
+
+        if start_image is not None:
+            start_image = comfy.utils.common_upscale(start_image[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
+
+            encoded = vae.encode(start_image[:, :, :, :3])
+            concat_latent_image = torch.zeros((latent.shape[0], 32, latent.shape[2], latent.shape[3], latent.shape[4]), device=comfy.model_management.intermediate_device())
+            concat_latent_image[:, :, :encoded.shape[2], :, :] = encoded
+
+            mask = torch.ones((1, 1, latent.shape[2], concat_latent_image.shape[-2], concat_latent_image.shape[-1]), device=start_image.device, dtype=start_image.dtype)
+            mask[:, :, :((start_image.shape[0] - 1) // 4) + 1] = 0.0
+
+            positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": concat_latent_image, "concat_mask": mask})
+            negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": concat_latent_image, "concat_mask": mask})
+
+        if clip_vision_output is not None:
+            positive = node_helpers.conditioning_set_values(positive, {"clip_vision_output": clip_vision_output})
+            negative = node_helpers.conditioning_set_values(negative, {"clip_vision_output": clip_vision_output})
+
+        out_latent = {}
+        out_latent["samples"] = latent
+        return io.NodeOutput(positive, negative, out_latent)
+
+
+class HunyuanVideo15SuperResolution(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="HunyuanVideo15SuperResolution",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae", optional=True),
+                io.Image.Input("start_image", optional=True),
+                io.ClipVisionOutput.Input("clip_vision_output", optional=True),
+                io.Latent.Input("latent"),
+                io.Float.Input("noise_augmentation", default=0.70, min=0.0, max=1.0, step=0.01),
+
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, positive, negative, latent, noise_augmentation, vae=None, start_image=None, clip_vision_output=None) -> io.NodeOutput:
+        in_latent = latent["samples"]
+        in_channels = in_latent.shape[1]
+        cond_latent = torch.zeros([in_latent.shape[0], in_channels * 2 + 2, in_latent.shape[-3], in_latent.shape[-2], in_latent.shape[-1]], device=comfy.model_management.intermediate_device())
+        cond_latent[:, in_channels + 1 : 2 * in_channels + 1] = in_latent
+        cond_latent[:, 2 * in_channels + 1] = 1
+        if start_image is not None:
+            start_image = comfy.utils.common_upscale(start_image.movedim(-1, 1), in_latent.shape[-1] * 16, in_latent.shape[-2] * 16, "bilinear", "center").movedim(1, -1)
+            encoded = vae.encode(start_image[:, :, :, :3])
+            cond_latent[:, :in_channels, :encoded.shape[2], :, :] = encoded
+            cond_latent[:, in_channels + 1, 0] = 1
+
+        positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": cond_latent, "noise_augmentation": noise_augmentation})
+        negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": cond_latent, "noise_augmentation": noise_augmentation})
+        if clip_vision_output is not None:
+            positive = node_helpers.conditioning_set_values(positive, {"clip_vision_output": clip_vision_output})
+            negative = node_helpers.conditioning_set_values(negative, {"clip_vision_output": clip_vision_output})
+
+        return io.NodeOutput(positive, negative, latent)
+
+
+class LatentUpscaleModelLoader(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="LatentUpscaleModelLoader",
+            display_name="Load Latent Upscale Model",
+            category="loaders",
+            inputs=[
+                io.Combo.Input("model_name", options=folder_paths.get_filename_list("latent_upscale_models")),
+            ],
+            outputs=[
+                io.LatentUpscaleModel.Output(),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, model_name) -> io.NodeOutput:
+        model_path = folder_paths.get_full_path_or_raise("latent_upscale_models", model_name)
+        sd = comfy.utils.load_torch_file(model_path, safe_load=True)
+
+        if "blocks.0.block.0.conv.weight" in sd:
+            config = {
+                "in_channels": sd["in_conv.conv.weight"].shape[1],
+                "out_channels": sd["out_conv.conv.weight"].shape[0],
+                "hidden_channels": sd["in_conv.conv.weight"].shape[0],
+                "num_blocks": len([k for k in sd.keys() if k.startswith("blocks.") and k.endswith(".block.0.conv.weight")]),
+                "global_residual": False,
+            }
+            model_type = "720p"
+        elif "up.0.block.0.conv1.conv.weight" in sd:
+            sd = {key.replace("nin_shortcut", "nin_shortcut.conv", 1): value for key, value in sd.items()}
+            config = {
+                "z_channels": sd["conv_in.conv.weight"].shape[1],
+                "out_channels": sd["conv_out.conv.weight"].shape[0],
+                "block_out_channels": tuple(sd[f"up.{i}.block.0.conv1.conv.weight"].shape[0] for i in range(len([k for k in sd.keys() if k.startswith("up.") and k.endswith(".block.0.conv1.conv.weight")]))),
+            }
+            model_type = "1080p"
+
+        model = HunyuanVideo15SRModel(model_type, config)
+        model.load_sd(sd)
+
+        return io.NodeOutput(model)
+
+
+class HunyuanVideo15LatentUpscaleWithModel(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="HunyuanVideo15LatentUpscaleWithModel",
+            display_name="Hunyuan Video 15 Latent Upscale With Model",
+            category="latent",
+            inputs=[
+                io.LatentUpscaleModel.Input("model"),
+                io.Latent.Input("samples"),
+                io.Combo.Input("upscale_method", options=["nearest-exact", "bilinear", "area", "bicubic", "bislerp"], default="bilinear"),
+                io.Int.Input("width", default=1280, min=0, max=16384, step=8),
+                io.Int.Input("height", default=720, min=0, max=16384, step=8),
+                io.Combo.Input("crop", options=["disabled", "center"]),
+            ],
+            outputs=[
+                io.Latent.Output(),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, model, samples, upscale_method, width, height, crop) -> io.NodeOutput:
+        if width == 0 and height == 0:
+            return io.NodeOutput(samples)
+        else:
+            if width == 0:
+                height = max(64, height)
+                width = max(64, round(samples["samples"].shape[-1] * height / samples["samples"].shape[-2]))
+            elif height == 0:
+                width = max(64, width)
+                height = max(64, round(samples["samples"].shape[-2] * width / samples["samples"].shape[-1]))
+            else:
+                width = max(64, width)
+                height = max(64, height)
+            s = comfy.utils.common_upscale(samples["samples"], width // 16, height // 16, upscale_method, crop)
+            s = model.resample_latent(s)
+            return io.NodeOutput({"samples": s.cpu().float()})
+
+
 PROMPT_TEMPLATE_ENCODE_VIDEO_I2V = (
     "<|start_header_id|>system<|end_header_id|>\n\n<image>\nDescribe the video by detailing the following aspects according to the reference image: "
     "1. The main content and theme of the video."
@@ -210,6 +404,11 @@ class HunyuanExtension(ComfyExtension):
             CLIPTextEncodeHunyuanDiT,
             TextEncodeHunyuanVideo_ImageToVideo,
             EmptyHunyuanLatentVideo,
+            EmptyHunyuanVideo15Latent,
+            HunyuanVideo15ImageToVideo,
+            HunyuanVideo15SuperResolution,
+            HunyuanVideo15LatentUpscaleWithModel,
+            LatentUpscaleModelLoader,
             HunyuanImageToVideo,
             EmptyHunyuanImageLatent,
             HunyuanRefinerLatent,

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.3.70"
+__version__ = "0.3.74"

folder_paths.py

@@ -38,6 +38,8 @@ folder_names_and_paths["gligen"] = ([os.path.join(models_dir, "gligen")], suppor
 
 folder_names_and_paths["upscale_models"] = ([os.path.join(models_dir, "upscale_models")], supported_pt_extensions)
 
+folder_names_and_paths["latent_upscale_models"] = ([os.path.join(models_dir, "latent_upscale_models")], supported_pt_extensions)
+
 folder_names_and_paths["custom_nodes"] = ([os.path.join(base_path, "custom_nodes")], set())
 
 folder_names_and_paths["hypernetworks"] = ([os.path.join(models_dir, "hypernetworks")], supported_pt_extensions)

nodes.py

@@ -929,7 +929,7 @@ class CLIPLoader:
     @classmethod
     def INPUT_TYPES(s):
         return {"required": { "clip_name": (folder_paths.get_filename_list("text_encoders"), ),
-                              "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream", "chroma", "ace", "omnigen2", "qwen_image", "hunyuan_image"], ),
+                              "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream", "chroma", "ace", "omnigen2", "qwen_image", "hunyuan_image", "flux2"], ),
                               },
                 "optional": {
                               "device": (["default", "cpu"], {"advanced": True}),
@@ -957,7 +957,7 @@ class DualCLIPLoader:
     def INPUT_TYPES(s):
         return {"required": { "clip_name1": (folder_paths.get_filename_list("text_encoders"), ),
                               "clip_name2": (folder_paths.get_filename_list("text_encoders"), ),
-                              "type": (["sdxl", "sd3", "flux", "hunyuan_video", "hidream", "hunyuan_image"], ),
+                              "type": (["sdxl", "sd3", "flux", "hunyuan_video", "hidream", "hunyuan_image", "hunyuan_video_15"], ),
                               },
                 "optional": {
                               "device": (["default", "cpu"], {"advanced": True}),

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.3.70"
+version = "0.3.74"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.9"

requirements.txt

@@ -1,5 +1,5 @@
-comfyui-frontend-package==1.28.9
-comfyui-workflow-templates==0.3.1
+comfyui-frontend-package==1.30.6
+comfyui-workflow-templates==0.7.20
 comfyui-embedded-docs==0.3.1
 torch
 torchsde
@@ -7,7 +7,7 @@ torchvision
 torchaudio
 numpy>=1.25.0
 einops
-transformers>=4.37.2
+transformers>=4.50.3
 tokenizers>=0.13.3
 sentencepiece
 safetensors>=0.4.2

server.py

@@ -164,6 +164,22 @@ def create_origin_only_middleware():
 
     return origin_only_middleware
 
+
+def create_block_external_middleware():
+    @web.middleware
+    async def block_external_middleware(request: web.Request, handler):
+        if request.method == "OPTIONS":
+            # Pre-flight request. Reply successfully:
+            response = web.Response()
+        else:
+            response = await handler(request)
+
+        response.headers['Content-Security-Policy'] = "default-src 'self'; script-src 'self' 'unsafe-inline' blob:; style-src 'self' 'unsafe-inline'; img-src 'self' data: blob:; font-src 'self'; connect-src 'self'; frame-src 'self'; object-src 'self';"
+        return response
+
+    return block_external_middleware
+
+
 class PromptServer():
     def __init__(self, loop):
         PromptServer.instance = self
@@ -193,6 +209,9 @@ class PromptServer():
         else:
             middlewares.append(create_origin_only_middleware())
 
+        if args.disable_api_nodes:
+            middlewares.append(create_block_external_middleware())
+
         max_upload_size = round(args.max_upload_size * 1024 * 1024)
         self.app = web.Application(client_max_size=max_upload_size, middlewares=middlewares)
         self.sockets = dict()

tests-unit/comfy_quant/test_mixed_precision.py

@@ -37,11 +37,8 @@ class TestMixedPrecisionOps(unittest.TestCase):
 
     def test_all_layers_standard(self):
         """Test that model with no quantization works normally"""
-        # Configure no quantization
-        ops.MixedPrecisionOps._layer_quant_config = {}
-
         # Create model
-        model = SimpleModel(operations=ops.MixedPrecisionOps)
+        model = SimpleModel(operations=ops.mixed_precision_ops({}))
 
         # Initialize weights manually
         model.layer1.weight = torch.nn.Parameter(torch.randn(20, 10, dtype=torch.bfloat16))
@@ -76,7 +73,6 @@ class TestMixedPrecisionOps(unittest.TestCase):
                 "params": {}
             }
         }
-        ops.MixedPrecisionOps._layer_quant_config = layer_quant_config
 
         # Create state dict with mixed precision
         fp8_weight1 = torch.randn(20, 10, dtype=torch.float32).to(torch.float8_e4m3fn)
@@ -99,7 +95,7 @@ class TestMixedPrecisionOps(unittest.TestCase):
         }
 
         # Create model and load state dict (strict=False because custom loading pops keys)
-        model = SimpleModel(operations=ops.MixedPrecisionOps)
+        model = SimpleModel(operations=ops.mixed_precision_ops(layer_quant_config))
         model.load_state_dict(state_dict, strict=False)
 
         # Verify weights are wrapped in QuantizedTensor
@@ -132,7 +128,6 @@ class TestMixedPrecisionOps(unittest.TestCase):
                 "params": {}
             }
         }
-        ops.MixedPrecisionOps._layer_quant_config = layer_quant_config
 
         # Create and load model
         fp8_weight = torch.randn(20, 10, dtype=torch.float32).to(torch.float8_e4m3fn)
@@ -146,7 +141,7 @@ class TestMixedPrecisionOps(unittest.TestCase):
             "layer3.bias": torch.randn(40, dtype=torch.bfloat16),
         }
 
-        model = SimpleModel(operations=ops.MixedPrecisionOps)
+        model = SimpleModel(operations=ops.mixed_precision_ops(layer_quant_config))
         model.load_state_dict(state_dict1, strict=False)
 
         # Save state dict
@@ -170,7 +165,6 @@ class TestMixedPrecisionOps(unittest.TestCase):
                 "params": {}
             }
         }
-        ops.MixedPrecisionOps._layer_quant_config = layer_quant_config
 
         # Create and load model
         fp8_weight = torch.randn(20, 10, dtype=torch.float32).to(torch.float8_e4m3fn)
@@ -184,7 +178,7 @@ class TestMixedPrecisionOps(unittest.TestCase):
             "layer3.bias": torch.randn(40, dtype=torch.bfloat16),
         }
 
-        model = SimpleModel(operations=ops.MixedPrecisionOps)
+        model = SimpleModel(operations=ops.mixed_precision_ops(layer_quant_config))
         model.load_state_dict(state_dict, strict=False)
 
         # Add a weight function (simulating LoRA)
@@ -210,7 +204,6 @@ class TestMixedPrecisionOps(unittest.TestCase):
                 "params": {}
             }
         }
-        ops.MixedPrecisionOps._layer_quant_config = layer_quant_config
 
         # Create state dict
         state_dict = {
@@ -223,7 +216,7 @@ class TestMixedPrecisionOps(unittest.TestCase):
         }
 
         # Load should raise KeyError for unknown format in QUANT_FORMAT_MIXINS
-        model = SimpleModel(operations=ops.MixedPrecisionOps)
+        model = SimpleModel(operations=ops.mixed_precision_ops(layer_quant_config))
         with self.assertRaises(KeyError):
             model.load_state_dict(state_dict, strict=False)
 

tests/execution/test_public_api.py

@@ -0,0 +1,153 @@
+"""
+Tests for public ComfyAPI and ComfyAPISync functions.
+
+These tests verify that the public API methods work correctly in both sync and async contexts,
+ensuring that the sync wrapper generation (via get_type_hints() in async_to_sync.py) correctly
+handles string annotations from 'from __future__ import annotations'.
+"""
+
+import pytest
+import time
+import subprocess
+import torch
+from pytest import fixture
+from comfy_execution.graph_utils import GraphBuilder
+from tests.execution.test_execution import ComfyClient
+
+
+@pytest.mark.execution
+class TestPublicAPI:
+    """Test suite for public ComfyAPI and ComfyAPISync methods."""
+
+    @fixture(scope="class", autouse=True)
+    def _server(self, args_pytest):
+        """Start ComfyUI server for testing."""
+        pargs = [
+            'python', 'main.py',
+            '--output-directory', args_pytest["output_dir"],
+            '--listen', args_pytest["listen"],
+            '--port', str(args_pytest["port"]),
+            '--extra-model-paths-config', 'tests/execution/extra_model_paths.yaml',
+            '--cpu',
+        ]
+        p = subprocess.Popen(pargs)
+        yield
+        p.kill()
+        torch.cuda.empty_cache()
+
+    @fixture(scope="class", autouse=True)
+    def shared_client(self, args_pytest, _server):
+        """Create shared client with connection retry."""
+        client = ComfyClient()
+        n_tries = 5
+        for i in range(n_tries):
+            time.sleep(4)
+            try:
+                client.connect(listen=args_pytest["listen"], port=args_pytest["port"])
+                break
+            except ConnectionRefusedError:
+                if i == n_tries - 1:
+                    raise
+        yield client
+        del client
+        torch.cuda.empty_cache()
+
+    @fixture
+    def client(self, shared_client, request):
+        """Set test name for each test."""
+        shared_client.set_test_name(f"public_api[{request.node.name}]")
+        yield shared_client
+
+    @fixture
+    def builder(self, request):
+        """Create GraphBuilder for each test."""
+        yield GraphBuilder(prefix=request.node.name)
+
+    def test_sync_progress_update_executes(self, client: ComfyClient, builder: GraphBuilder):
+        """Test that TestSyncProgressUpdate executes without errors.
+
+        This test validates that api_sync.execution.set_progress() works correctly,
+        which is the primary code path fixed by adding get_type_hints() to async_to_sync.py.
+        """
+        g = builder
+        image = g.node("StubImage", content="BLACK", height=256, width=256, batch_size=1)
+
+        # Use TestSyncProgressUpdate with short sleep
+        progress_node = g.node("TestSyncProgressUpdate",
+                              value=image.out(0),
+                              sleep_seconds=0.5)
+        output = g.node("SaveImage", images=progress_node.out(0))
+
+        # Execute workflow
+        result = client.run(g)
+
+        # Verify execution
+        assert result.did_run(progress_node), "Progress node should have executed"
+        assert result.did_run(output), "Output node should have executed"
+
+        # Verify output
+        images = result.get_images(output)
+        assert len(images) == 1, "Should have produced 1 image"
+
+    def test_async_progress_update_executes(self, client: ComfyClient, builder: GraphBuilder):
+        """Test that TestAsyncProgressUpdate executes without errors.
+
+        This test validates that await api.execution.set_progress() works correctly
+        in async contexts.
+        """
+        g = builder
+        image = g.node("StubImage", content="WHITE", height=256, width=256, batch_size=1)
+
+        # Use TestAsyncProgressUpdate with short sleep
+        progress_node = g.node("TestAsyncProgressUpdate",
+                              value=image.out(0),
+                              sleep_seconds=0.5)
+        output = g.node("SaveImage", images=progress_node.out(0))
+
+        # Execute workflow
+        result = client.run(g)
+
+        # Verify execution
+        assert result.did_run(progress_node), "Async progress node should have executed"
+        assert result.did_run(output), "Output node should have executed"
+
+        # Verify output
+        images = result.get_images(output)
+        assert len(images) == 1, "Should have produced 1 image"
+
+    def test_sync_and_async_progress_together(self, client: ComfyClient, builder: GraphBuilder):
+        """Test both sync and async progress updates in same workflow.
+
+        This test ensures that both ComfyAPISync and ComfyAPI can coexist and work
+        correctly in the same workflow execution.
+        """
+        g = builder
+        image1 = g.node("StubImage", content="BLACK", height=256, width=256, batch_size=1)
+        image2 = g.node("StubImage", content="WHITE", height=256, width=256, batch_size=1)
+
+        # Use both types of progress nodes
+        sync_progress = g.node("TestSyncProgressUpdate",
+                              value=image1.out(0),
+                              sleep_seconds=0.3)
+        async_progress = g.node("TestAsyncProgressUpdate",
+                               value=image2.out(0),
+                               sleep_seconds=0.3)
+
+        # Create outputs
+        output1 = g.node("SaveImage", images=sync_progress.out(0))
+        output2 = g.node("SaveImage", images=async_progress.out(0))
+
+        # Execute workflow
+        result = client.run(g)
+
+        # Both should execute successfully
+        assert result.did_run(sync_progress), "Sync progress node should have executed"
+        assert result.did_run(async_progress), "Async progress node should have executed"
+        assert result.did_run(output1), "First output node should have executed"
+        assert result.did_run(output2), "Second output node should have executed"
+
+        # Verify outputs
+        images1 = result.get_images(output1)
+        images2 = result.get_images(output2)
+        assert len(images1) == 1, "Should have produced 1 image from sync node"
+        assert len(images2) == 1, "Should have produced 1 image from async node"