Fix canny node not working with fp16. (#13085)

The recent PR that added resize_cond_for_context_window methods to
model classes used inline 'import comfy.context_windows' in each
method body. This moves that import to the top-level import section,
replacing 4 duplicate inline imports with a single top-level one.

comfy/context_windows.py

@@ -93,6 +93,50 @@ class IndexListCallbacks:
         return {}
 
 
+def slice_cond(cond_value, window: IndexListContextWindow, x_in: torch.Tensor, device, temporal_dim: int, temporal_scale: int=1, temporal_offset: int=0, retain_index_list: list[int]=[]):
+    if not (hasattr(cond_value, "cond") and isinstance(cond_value.cond, torch.Tensor)):
+        return None
+    cond_tensor = cond_value.cond
+    if temporal_dim >= cond_tensor.ndim:
+        return None
+
+    cond_size = cond_tensor.size(temporal_dim)
+
+    if temporal_scale == 1:
+        expected_size = x_in.size(window.dim) - temporal_offset
+        if cond_size != expected_size:
+            return None
+
+    if temporal_offset == 0 and temporal_scale == 1:
+        sliced = window.get_tensor(cond_tensor, device, dim=temporal_dim, retain_index_list=retain_index_list)
+        return cond_value._copy_with(sliced)
+
+    # skip leading latent positions that have no corresponding conditioning (e.g. reference frames)
+    if temporal_offset > 0:
+        indices = [i - temporal_offset for i in window.index_list[temporal_offset:]]
+        indices = [i for i in indices if 0 <= i]
+    else:
+        indices = list(window.index_list)
+
+    if not indices:
+        return None
+
+    if temporal_scale > 1:
+        scaled = []
+        for i in indices:
+            for k in range(temporal_scale):
+                si = i * temporal_scale + k
+                if si < cond_size:
+                    scaled.append(si)
+        indices = scaled
+        if not indices:
+            return None
+
+    idx = tuple([slice(None)] * temporal_dim + [indices])
+    sliced = cond_tensor[idx].to(device)
+    return cond_value._copy_with(sliced)
+
+
 @dataclass
 class ContextSchedule:
     name: str
@@ -177,10 +221,17 @@ class IndexListContextHandler(ContextHandlerABC):
                                     new_cond_item[cond_key] = result
                                     handled = True
                                     break
+                            if not handled and self._model is not None:
+                                result = self._model.resize_cond_for_context_window(
+                                    cond_key, cond_value, window, x_in, device,
+                                    retain_index_list=self.cond_retain_index_list)
+                                if result is not None:
+                                    new_cond_item[cond_key] = result
+                                    handled = True
                             if handled:
                                 continue
                             if isinstance(cond_value, torch.Tensor):
-                                if (self.dim < cond_value.ndim and cond_value(self.dim) == x_in.size(self.dim)) or \
+                                if (self.dim < cond_value.ndim and cond_value.size(self.dim) == x_in.size(self.dim)) or \
                                    (cond_value.ndim < self.dim and cond_value.size(0) == x_in.size(self.dim)):
                                     new_cond_item[cond_key] = window.get_tensor(cond_value, device)
                             # Handle audio_embed (temporal dim is 1)
@@ -224,6 +275,7 @@ class IndexListContextHandler(ContextHandlerABC):
         return context_windows
 
     def execute(self, calc_cond_batch: Callable, model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]):
+        self._model = model
         self.set_step(timestep, model_options)
         context_windows = self.get_context_windows(model, x_in, model_options)
         enumerated_context_windows = list(enumerate(context_windows))

comfy/ldm/lightricks/vae/causal_conv3d.py

@@ -23,6 +23,11 @@ class CausalConv3d(nn.Module):
         self.in_channels = in_channels
         self.out_channels = out_channels
 
+        if isinstance(stride, int):
+            self.time_stride = stride
+        else:
+            self.time_stride = stride[0]
+
         kernel_size = (kernel_size, kernel_size, kernel_size)
         self.time_kernel_size = kernel_size[0]
 
@@ -58,18 +63,23 @@ class CausalConv3d(nn.Module):
         pieces = [ cached, x ]
         if is_end and not causal:
             pieces.append(x[:, :, -1:, :, :].repeat((1, 1, (self.time_kernel_size - 1) // 2, 1, 1)))
+        input_length = sum([piece.shape[2] for piece in pieces])
+        cache_length = (self.time_kernel_size - self.time_stride) + ((input_length - self.time_kernel_size) % self.time_stride)
 
         needs_caching = not is_end
-        if needs_caching and x.shape[2] >= self.time_kernel_size - 1:
+        if needs_caching and cache_length == 0:
+            self.temporal_cache_state[tid] = (x[:, :, :0, :, :], False)
             needs_caching = False
-            self.temporal_cache_state[tid] = (x[:, :, -(self.time_kernel_size - 1):, :, :], False)
+        if needs_caching and x.shape[2] >= cache_length:
+            needs_caching = False
+            self.temporal_cache_state[tid] = (x[:, :, -cache_length:, :, :], False)
 
         x = torch.cat(pieces, dim=2)
         del pieces
         del cached
 
         if needs_caching:
-            self.temporal_cache_state[tid] = (x[:, :, -(self.time_kernel_size - 1):, :, :], False)
+            self.temporal_cache_state[tid] = (x[:, :, -cache_length:, :, :], False)
         elif is_end:
             self.temporal_cache_state[tid] = (None, True)
 

comfy/ldm/lightricks/vae/causal_video_autoencoder.py

@@ -233,10 +233,7 @@ class Encoder(nn.Module):
 
         self.gradient_checkpointing = False
 
-    def forward_orig(self, sample: torch.FloatTensor) -> torch.FloatTensor:
-        r"""The forward method of the `Encoder` class."""
-
-        sample = patchify(sample, patch_size_hw=self.patch_size, patch_size_t=1)
+    def _forward_chunk(self, sample: torch.FloatTensor) -> Optional[torch.FloatTensor]:
         sample = self.conv_in(sample)
 
         checkpoint_fn = (
@@ -247,10 +244,14 @@ class Encoder(nn.Module):
 
         for down_block in self.down_blocks:
             sample = checkpoint_fn(down_block)(sample)
+            if sample is None or sample.shape[2] == 0:
+                return None
 
         sample = self.conv_norm_out(sample)
         sample = self.conv_act(sample)
         sample = self.conv_out(sample)
+        if sample is None or sample.shape[2] == 0:
+            return None
 
         if self.latent_log_var == "uniform":
             last_channel = sample[:, -1:, ...]
@@ -282,9 +283,35 @@ class Encoder(nn.Module):
 
         return sample
 
+    def forward_orig(self, sample: torch.FloatTensor, device=None) -> torch.FloatTensor:
+        r"""The forward method of the `Encoder` class."""
+
+        max_chunk_size = get_max_chunk_size(sample.device if device is None else device) * 2  # encoder is more memory-efficient than decoder
+        frame_size = sample[:, :, :1, :, :].numel() * sample.element_size()
+        frame_size = int(frame_size * (self.conv_in.out_channels / self.conv_in.in_channels))
+
+        outputs = []
+        samples = [sample[:, :, :1, :, :]]
+        if sample.shape[2] > 1:
+            chunk_t = max(2, max_chunk_size // frame_size)
+            if chunk_t < 4:
+                chunk_t = 2
+            elif chunk_t < 8:
+                chunk_t = 4
+            else:
+                chunk_t = (chunk_t // 8) * 8
+            samples += list(torch.split(sample[:, :, 1:, :, :], chunk_t, dim=2))
+        for chunk_idx, chunk in enumerate(samples):
+            if chunk_idx == len(samples) - 1:
+                mark_conv3d_ended(self)
+            chunk = patchify(chunk, patch_size_hw=self.patch_size, patch_size_t=1).to(device=device)
+            output = self._forward_chunk(chunk)
+            if output is not None:
+                outputs.append(output)
+
+        return torch_cat_if_needed(outputs, dim=2)
+
     def forward(self, *args, **kwargs):
-        #No encoder support so just flag the end so it doesnt use the cache.
-        mark_conv3d_ended(self)
         try:
             return self.forward_orig(*args, **kwargs)
         finally:
@@ -473,6 +500,17 @@ class Decoder(nn.Module):
 
         self.gradient_checkpointing = False
 
+        # Precompute output scale factors: (channels, (t_scale, h_scale, w_scale), t_offset)
+        ts, hs, ws, to = 1, 1, 1, 0
+        for block in self.up_blocks:
+            if isinstance(block, DepthToSpaceUpsample):
+                ts *= block.stride[0]
+                hs *= block.stride[1]
+                ws *= block.stride[2]
+                if block.stride[0] > 1:
+                    to = to * block.stride[0] + 1
+        self._output_scale = (out_channels // (patch_size ** 2), (ts, hs * patch_size, ws * patch_size), to)
+
         self.timestep_conditioning = timestep_conditioning
 
         if timestep_conditioning:
@@ -494,11 +532,62 @@ class Decoder(nn.Module):
             )
 
 
-    # def forward(self, sample: torch.FloatTensor, target_shape) -> torch.FloatTensor:
+    def decode_output_shape(self, input_shape):
+        c, (ts, hs, ws), to = self._output_scale
+        return (input_shape[0], c, input_shape[2] * ts - to, input_shape[3] * hs, input_shape[4] * ws)
+
+    def run_up(self, idx, sample_ref, ended, timestep_shift_scale, scaled_timestep, checkpoint_fn, output_buffer, output_offset, max_chunk_size):
+        sample = sample_ref[0]
+        sample_ref[0] = None
+        if idx >= len(self.up_blocks):
+            sample = self.conv_norm_out(sample)
+            if timestep_shift_scale is not None:
+                shift, scale = timestep_shift_scale
+                sample = sample * (1 + scale) + shift
+            sample = self.conv_act(sample)
+            if ended:
+                mark_conv3d_ended(self.conv_out)
+            sample = self.conv_out(sample, causal=self.causal)
+            if sample is not None and sample.shape[2] > 0:
+                sample = unpatchify(sample, patch_size_hw=self.patch_size, patch_size_t=1)
+                t = sample.shape[2]
+                output_buffer[:, :, output_offset[0]:output_offset[0] + t].copy_(sample)
+                output_offset[0] += t
+            return
+
+        up_block = self.up_blocks[idx]
+        if ended:
+            mark_conv3d_ended(up_block)
+        if self.timestep_conditioning and isinstance(up_block, UNetMidBlock3D):
+            sample = checkpoint_fn(up_block)(
+                sample, causal=self.causal, timestep=scaled_timestep
+            )
+        else:
+            sample = checkpoint_fn(up_block)(sample, causal=self.causal)
+
+        if sample is None or sample.shape[2] == 0:
+            return
+
+        total_bytes = sample.numel() * sample.element_size()
+        num_chunks = (total_bytes + max_chunk_size - 1) // max_chunk_size
+
+        if num_chunks == 1:
+            # when we are not chunking, detach our x so the callee can free it as soon as they are done
+            next_sample_ref = [sample]
+            del sample
+            self.run_up(idx + 1, next_sample_ref, ended, timestep_shift_scale, scaled_timestep, checkpoint_fn, output_buffer, output_offset, max_chunk_size)
+            return
+        else:
+            samples = torch.chunk(sample, chunks=num_chunks, dim=2)
+
+            for chunk_idx, sample1 in enumerate(samples):
+                self.run_up(idx + 1, [sample1], ended and chunk_idx == len(samples) - 1, timestep_shift_scale, scaled_timestep, checkpoint_fn, output_buffer, output_offset, max_chunk_size)
+
     def forward_orig(
         self,
         sample: torch.FloatTensor,
         timestep: Optional[torch.Tensor] = None,
+        output_buffer: Optional[torch.Tensor] = None,
     ) -> torch.FloatTensor:
         r"""The forward method of the `Decoder` class."""
         batch_size = sample.shape[0]
@@ -513,6 +602,7 @@ class Decoder(nn.Module):
         )
 
         timestep_shift_scale = None
+        scaled_timestep = None
         if self.timestep_conditioning:
             assert (
                 timestep is not None
@@ -540,59 +630,18 @@ class Decoder(nn.Module):
             )
             timestep_shift_scale = ada_values.unbind(dim=1)
 
-        output = []
+        if output_buffer is None:
+            output_buffer = torch.empty(
+                self.decode_output_shape(sample.shape),
+                dtype=sample.dtype, device=comfy.model_management.intermediate_device(),
+            )
+        output_offset = [0]
+
         max_chunk_size = get_max_chunk_size(sample.device)
 
-        def run_up(idx, sample_ref, ended):
-            sample = sample_ref[0]
-            sample_ref[0] = None
-            if idx >= len(self.up_blocks):
-                sample = self.conv_norm_out(sample)
-                if timestep_shift_scale is not None:
-                    shift, scale = timestep_shift_scale
-                    sample = sample * (1 + scale) + shift
-                sample = self.conv_act(sample)
-                if ended:
-                    mark_conv3d_ended(self.conv_out)
-                sample = self.conv_out(sample, causal=self.causal)
-                if sample is not None and sample.shape[2] > 0:
-                    output.append(sample.to(comfy.model_management.intermediate_device()))
-                return
+        self.run_up(0, [sample], True, timestep_shift_scale, scaled_timestep, checkpoint_fn, output_buffer, output_offset, max_chunk_size)
 
-            up_block = self.up_blocks[idx]
-            if (ended):
-                mark_conv3d_ended(up_block)
-            if self.timestep_conditioning and isinstance(up_block, UNetMidBlock3D):
-                sample = checkpoint_fn(up_block)(
-                    sample, causal=self.causal, timestep=scaled_timestep
-                )
-            else:
-                sample = checkpoint_fn(up_block)(sample, causal=self.causal)
-
-            if sample is None or sample.shape[2] == 0:
-                return
-
-            total_bytes = sample.numel() * sample.element_size()
-            num_chunks = (total_bytes + max_chunk_size - 1) // max_chunk_size
-
-            if num_chunks == 1:
-                # when we are not chunking, detach our x so the callee can free it as soon as they are done
-                next_sample_ref = [sample]
-                del sample
-                run_up(idx + 1, next_sample_ref, ended)
-                return
-            else:
-                samples = torch.chunk(sample, chunks=num_chunks, dim=2)
-
-                for chunk_idx, sample1 in enumerate(samples):
-                    run_up(idx + 1, [sample1], ended and chunk_idx == len(samples) - 1)
-
-        run_up(0, [sample], True)
-        sample = torch.cat(output, dim=2)
-
-        sample = unpatchify(sample, patch_size_hw=self.patch_size, patch_size_t=1)
-
-        return sample
+        return output_buffer
 
     def forward(self, *args, **kwargs):
         try:
@@ -716,12 +765,25 @@ class SpaceToDepthDownsample(nn.Module):
             causal=True,
             spatial_padding_mode=spatial_padding_mode,
         )
+        self.temporal_cache_state = {}
 
     def forward(self, x, causal: bool = True):
-        if self.stride[0] == 2:
+        tid = threading.get_ident()
+        cached, pad_first, cached_x, cached_input = self.temporal_cache_state.get(tid, (None, True, None, None))
+        if cached_input is not None:
+            x = torch_cat_if_needed([cached_input, x], dim=2)
+            cached_input = None
+
+        if self.stride[0] == 2 and pad_first:
             x = torch.cat(
                 [x[:, :, :1, :, :], x], dim=2
             )  # duplicate first frames for padding
+            pad_first = False
+
+        if x.shape[2] < self.stride[0]:
+            cached_input = x
+            self.temporal_cache_state[tid] = (cached, pad_first, cached_x, cached_input)
+            return None
 
         # skip connection
         x_in = rearrange(
@@ -736,15 +798,26 @@ class SpaceToDepthDownsample(nn.Module):
 
         # conv
         x = self.conv(x, causal=causal)
-        x = rearrange(
-            x,
-            "b c (d p1) (h p2) (w p3) -> b (c p1 p2 p3) d h w",
-            p1=self.stride[0],
-            p2=self.stride[1],
-            p3=self.stride[2],
-        )
+        if self.stride[0] == 2 and x.shape[2] == 1:
+            if cached_x is not None:
+                x = torch_cat_if_needed([cached_x, x], dim=2)
+                cached_x = None
+            else:
+                cached_x = x
+                x = None
 
-        x = x + x_in
+        if x is not None:
+            x = rearrange(
+                x,
+                "b c (d p1) (h p2) (w p3) -> b (c p1 p2 p3) d h w",
+                p1=self.stride[0],
+                p2=self.stride[1],
+                p3=self.stride[2],
+            )
+
+        cached = add_exchange_cache(x, cached, x_in, dim=2)
+
+        self.temporal_cache_state[tid] = (cached, pad_first, cached_x, cached_input)
 
         return x
 
@@ -1077,6 +1150,8 @@ class processor(nn.Module):
         return (x - self.get_buffer("mean-of-means").view(1, -1, 1, 1, 1).to(x)) / self.get_buffer("std-of-means").view(1, -1, 1, 1, 1).to(x)
 
 class VideoVAE(nn.Module):
+    comfy_has_chunked_io = True
+
     def __init__(self, version=0, config=None):
         super().__init__()
 
@@ -1219,14 +1294,15 @@ class VideoVAE(nn.Module):
             }
         return config
 
-    def encode(self, x):
-        frames_count = x.shape[2]
-        if ((frames_count - 1) % 8) != 0:
-            raise ValueError("Invalid number of frames: Encode input must have 1 + 8 * x frames (e.g., 1, 9, 17, ...). Please check your input.")
-        means, logvar = torch.chunk(self.encoder(x), 2, dim=1)
+    def encode(self, x, device=None):
+        x = x[:, :, :max(1, 1 + ((x.shape[2] - 1) // 8) * 8), :, :]
+        means, logvar = torch.chunk(self.encoder(x, device=device), 2, dim=1)
         return self.per_channel_statistics.normalize(means)
 
-    def decode(self, x):
+    def decode_output_shape(self, input_shape):
+        return self.decoder.decode_output_shape(input_shape)
+
+    def decode(self, x, output_buffer=None):
         if self.timestep_conditioning: #TODO: seed
             x = torch.randn_like(x) * self.decode_noise_scale + (1.0 - self.decode_noise_scale) * x
-        return self.decoder(self.per_channel_statistics.un_normalize(x), timestep=self.decode_timestep)
+        return self.decoder(self.per_channel_statistics.un_normalize(x), timestep=self.decode_timestep, output_buffer=output_buffer)

comfy/ldm/wan/vae.py

@@ -360,6 +360,43 @@ class Decoder3d(nn.Module):
             RMS_norm(out_dim, images=False), nn.SiLU(),
             CausalConv3d(out_dim, output_channels, 3, padding=1))
 
+    def run_up(self, layer_idx, x_ref, feat_cache, feat_idx, out_chunks):
+        x = x_ref[0]
+        x_ref[0] = None
+        if layer_idx >= len(self.upsamples):
+            for layer in self.head:
+                if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                    cache_x = x[:, :, -CACHE_T:, :, :]
+                    x = layer(x, feat_cache[feat_idx[0]])
+                    feat_cache[feat_idx[0]] = cache_x
+                    feat_idx[0] += 1
+                else:
+                    x = layer(x)
+            out_chunks.append(x)
+            return
+
+        layer = self.upsamples[layer_idx]
+        if isinstance(layer, Resample) and layer.mode == 'upsample3d' and x.shape[2] > 1:
+            for frame_idx in range(x.shape[2]):
+                self.run_up(
+                    layer_idx,
+                    [x[:, :, frame_idx:frame_idx + 1, :, :]],
+                    feat_cache,
+                    feat_idx.copy(),
+                    out_chunks,
+                )
+            del x
+            return
+
+        if feat_cache is not None:
+            x = layer(x, feat_cache, feat_idx)
+        else:
+            x = layer(x)
+
+        next_x_ref = [x]
+        del x
+        self.run_up(layer_idx + 1, next_x_ref, feat_cache, feat_idx, out_chunks)
+
     def forward(self, x, feat_cache=None, feat_idx=[0]):
         ## conv1
         if feat_cache is not None:
@@ -380,42 +417,7 @@ class Decoder3d(nn.Module):
 
         out_chunks = []
 
-        def run_up(layer_idx, x_ref, feat_idx):
-            x = x_ref[0]
-            x_ref[0] = None
-            if layer_idx >= len(self.upsamples):
-                for layer in self.head:
-                    if isinstance(layer, CausalConv3d) and feat_cache is not None:
-                        cache_x = x[:, :, -CACHE_T:, :, :]
-                        x = layer(x, feat_cache[feat_idx[0]])
-                        feat_cache[feat_idx[0]] = cache_x
-                        feat_idx[0] += 1
-                    else:
-                        x = layer(x)
-                out_chunks.append(x)
-                return
-
-            layer = self.upsamples[layer_idx]
-            if isinstance(layer, Resample) and layer.mode == 'upsample3d' and x.shape[2] > 1:
-                for frame_idx in range(x.shape[2]):
-                    run_up(
-                        layer_idx,
-                        [x[:, :, frame_idx:frame_idx + 1, :, :]],
-                        feat_idx.copy(),
-                    )
-                del x
-                return
-
-            if feat_cache is not None:
-                x = layer(x, feat_cache, feat_idx)
-            else:
-                x = layer(x)
-
-            next_x_ref = [x]
-            del x
-            run_up(layer_idx + 1, next_x_ref, feat_idx)
-
-        run_up(0, [x], feat_idx)
+        self.run_up(0, [x], feat_cache, feat_idx, out_chunks)
         return out_chunks
 
 

comfy/memory_management.py

@@ -39,7 +39,10 @@ def read_tensor_file_slice_into(tensor, destination):
     if (destination.device.type != "cpu"
             or file_obj is None
             or threading.get_ident() != info.thread_id
-            or destination.numel() * destination.element_size() < info.size):
+            or destination.numel() * destination.element_size() < info.size
+            or tensor.numel() * tensor.element_size() != info.size
+            or tensor.storage_offset() != 0
+            or not tensor.is_contiguous()):
         return False
 
     if info.size == 0:

comfy/model_base.py

@@ -21,6 +21,7 @@ import comfy.ldm.hunyuan3dv2_1.hunyuandit
 import torch
 import logging
 import comfy.ldm.lightricks.av_model
+import comfy.context_windows
 from comfy.ldm.modules.diffusionmodules.openaimodel import UNetModel, Timestep
 from comfy.ldm.cascade.stage_c import StageC
 from comfy.ldm.cascade.stage_b import StageB
@@ -285,6 +286,12 @@ class BaseModel(torch.nn.Module):
             return data
         return None
 
+    def resize_cond_for_context_window(self, cond_key, cond_value, window, x_in, device, retain_index_list=[]):
+        """Override in subclasses to handle model-specific cond slicing for context windows.
+        Return a sliced cond object, or None to fall through to default handling.
+        Use comfy.context_windows.slice_cond() for common cases."""
+        return None
+
     def extra_conds(self, **kwargs):
         out = {}
         concat_cond = self.concat_cond(**kwargs)
@@ -1375,6 +1382,11 @@ class WAN21_Vace(WAN21):
         out['vace_strength'] = comfy.conds.CONDConstant(vace_strength)
         return out
 
+    def resize_cond_for_context_window(self, cond_key, cond_value, window, x_in, device, retain_index_list=[]):
+        if cond_key == "vace_context":
+            return comfy.context_windows.slice_cond(cond_value, window, x_in, device, temporal_dim=3, retain_index_list=retain_index_list)
+        return super().resize_cond_for_context_window(cond_key, cond_value, window, x_in, device, retain_index_list=retain_index_list)
+
 class WAN21_Camera(WAN21):
     def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
         super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.CameraWanModel)
@@ -1427,6 +1439,11 @@ class WAN21_HuMo(WAN21):
 
         return out
 
+    def resize_cond_for_context_window(self, cond_key, cond_value, window, x_in, device, retain_index_list=[]):
+        if cond_key == "audio_embed":
+            return comfy.context_windows.slice_cond(cond_value, window, x_in, device, temporal_dim=1)
+        return super().resize_cond_for_context_window(cond_key, cond_value, window, x_in, device, retain_index_list=retain_index_list)
+
 class WAN22_Animate(WAN21):
     def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
         super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model_animate.AnimateWanModel)
@@ -1444,6 +1461,13 @@ class WAN22_Animate(WAN21):
             out['pose_latents'] = comfy.conds.CONDRegular(self.process_latent_in(pose_latents))
         return out
 
+    def resize_cond_for_context_window(self, cond_key, cond_value, window, x_in, device, retain_index_list=[]):
+        if cond_key == "face_pixel_values":
+            return comfy.context_windows.slice_cond(cond_value, window, x_in, device, temporal_dim=2, temporal_scale=4, temporal_offset=1)
+        if cond_key == "pose_latents":
+            return comfy.context_windows.slice_cond(cond_value, window, x_in, device, temporal_dim=2, temporal_offset=1)
+        return super().resize_cond_for_context_window(cond_key, cond_value, window, x_in, device, retain_index_list=retain_index_list)
+
 class WAN22_S2V(WAN21):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel_S2V)
@@ -1480,6 +1504,11 @@ class WAN22_S2V(WAN21):
             out['reference_motion'] = reference_motion.shape
         return out
 
+    def resize_cond_for_context_window(self, cond_key, cond_value, window, x_in, device, retain_index_list=[]):
+        if cond_key == "audio_embed":
+            return comfy.context_windows.slice_cond(cond_value, window, x_in, device, temporal_dim=1)
+        return super().resize_cond_for_context_window(cond_key, cond_value, window, x_in, device, retain_index_list=retain_index_list)
+
 class WAN22(WAN21):
     def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
         super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel)

comfy/model_management.py

@@ -1003,7 +1003,7 @@ def text_encoder_offload_device():
 def text_encoder_device():
     if args.gpu_only:
         return get_torch_device()
-    elif vram_state in (VRAMState.HIGH_VRAM, VRAMState.NORMAL_VRAM, VRAMState.SHARED) or comfy.memory_management.aimdo_enabled:
+    elif vram_state in (VRAMState.HIGH_VRAM, VRAMState.NORMAL_VRAM) or comfy.memory_management.aimdo_enabled:
         if should_use_fp16(prioritize_performance=False):
             return get_torch_device()
         else:

comfy/sample.py

@@ -64,10 +64,10 @@ def sample(model, noise, steps, cfg, sampler_name, scheduler, positive, negative
     sampler = comfy.samplers.KSampler(model, steps=steps, device=model.load_device, sampler=sampler_name, scheduler=scheduler, denoise=denoise, model_options=model.model_options)
 
     samples = sampler.sample(noise, positive, negative, cfg=cfg, latent_image=latent_image, start_step=start_step, last_step=last_step, force_full_denoise=force_full_denoise, denoise_mask=noise_mask, sigmas=sigmas, callback=callback, disable_pbar=disable_pbar, seed=seed)
-    samples = samples.to(comfy.model_management.intermediate_device())
+    samples = samples.to(device=comfy.model_management.intermediate_device(), dtype=comfy.model_management.intermediate_dtype())
     return samples
 
 def sample_custom(model, noise, cfg, sampler, sigmas, positive, negative, latent_image, noise_mask=None, callback=None, disable_pbar=False, seed=None):
     samples = comfy.samplers.sample(model, noise, positive, negative, cfg, model.load_device, sampler, sigmas, model_options=model.model_options, latent_image=latent_image, denoise_mask=noise_mask, callback=callback, disable_pbar=disable_pbar, seed=seed)
-    samples = samples.to(comfy.model_management.intermediate_device())
+    samples = samples.to(device=comfy.model_management.intermediate_device(), dtype=comfy.model_management.intermediate_dtype())
     return samples

comfy/sd.py

@@ -951,12 +951,23 @@ class VAE:
             batch_number = int(free_memory / memory_used)
             batch_number = max(1, batch_number)
 
+            # Pre-allocate output for VAEs that support direct buffer writes
+            preallocated = False
+            if getattr(self.first_stage_model, 'comfy_has_chunked_io', False):
+                pixel_samples = torch.empty(self.first_stage_model.decode_output_shape(samples_in.shape), device=self.output_device, dtype=self.vae_output_dtype())
+                preallocated = True
+
             for x in range(0, samples_in.shape[0], batch_number):
                 samples = samples_in[x:x + batch_number].to(device=self.device, dtype=self.vae_dtype)
-                out = self.process_output(self.first_stage_model.decode(samples, **vae_options).to(device=self.output_device, dtype=self.vae_output_dtype(), copy=True))
-                if pixel_samples is None:
-                    pixel_samples = torch.empty((samples_in.shape[0],) + tuple(out.shape[1:]), device=self.output_device, dtype=self.vae_output_dtype())
-                pixel_samples[x:x+batch_number] = out
+                if preallocated:
+                    self.first_stage_model.decode(samples, output_buffer=pixel_samples[x:x+batch_number], **vae_options)
+                else:
+                    out = self.first_stage_model.decode(samples, **vae_options).to(device=self.output_device, dtype=self.vae_output_dtype(), copy=True)
+                    if pixel_samples is None:
+                        pixel_samples = torch.empty((samples_in.shape[0],) + tuple(out.shape[1:]), device=self.output_device, dtype=self.vae_output_dtype())
+                    pixel_samples[x:x+batch_number].copy_(out)
+                    del out
+                self.process_output(pixel_samples[x:x+batch_number])
         except Exception as e:
             model_management.raise_non_oom(e)
             logging.warning("Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.")
@@ -967,6 +978,7 @@ class VAE:
             do_tile = True
 
         if do_tile:
+            comfy.model_management.soft_empty_cache()
             dims = samples_in.ndim - 2
             if dims == 1 or self.extra_1d_channel is not None:
                 pixel_samples = self.decode_tiled_1d(samples_in)
@@ -1027,8 +1039,13 @@ class VAE:
             batch_number = max(1, batch_number)
             samples = None
             for x in range(0, pixel_samples.shape[0], batch_number):
-                pixels_in = self.process_input(pixel_samples[x:x + batch_number]).to(self.vae_dtype).to(self.device)
-                out = self.first_stage_model.encode(pixels_in).to(self.output_device).to(dtype=self.vae_output_dtype())
+                pixels_in = self.process_input(pixel_samples[x:x + batch_number]).to(self.vae_dtype)
+                if getattr(self.first_stage_model, 'comfy_has_chunked_io', False):
+                    out = self.first_stage_model.encode(pixels_in, device=self.device)
+                else:
+                    pixels_in = pixels_in.to(self.device)
+                    out = self.first_stage_model.encode(pixels_in)
+                out = out.to(self.output_device).to(dtype=self.vae_output_dtype())
                 if samples is None:
                     samples = torch.empty((pixel_samples.shape[0],) + tuple(out.shape[1:]), device=self.output_device, dtype=self.vae_output_dtype())
                 samples[x:x + batch_number] = out
@@ -1043,6 +1060,7 @@ class VAE:
             do_tile = True
 
         if do_tile:
+            comfy.model_management.soft_empty_cache()
             if self.latent_dim == 3:
                 tile = 256
                 overlap = tile // 4

comfy/sd1_clip.py

@@ -46,7 +46,7 @@ class ClipTokenWeightEncoder:
         out, pooled = o[:2]
 
         if pooled is not None:
-            first_pooled = pooled[0:1].to(model_management.intermediate_device())
+            first_pooled = pooled[0:1].to(device=model_management.intermediate_device())
         else:
             first_pooled = pooled
 
@@ -63,16 +63,16 @@ class ClipTokenWeightEncoder:
             output.append(z)
 
         if (len(output) == 0):
-            r = (out[-1:].to(model_management.intermediate_device()), first_pooled)
+            r = (out[-1:].to(device=model_management.intermediate_device()), first_pooled)
         else:
-            r = (torch.cat(output, dim=-2).to(model_management.intermediate_device()), first_pooled)
+            r = (torch.cat(output, dim=-2).to(device=model_management.intermediate_device()), first_pooled)
 
         if len(o) > 2:
             extra = {}
             for k in o[2]:
                 v = o[2][k]
                 if k == "attention_mask":
-                    v = v[:sections].flatten().unsqueeze(dim=0).to(model_management.intermediate_device())
+                    v = v[:sections].flatten().unsqueeze(dim=0).to(device=model_management.intermediate_device())
                 extra[k] = v
 
             r = r + (extra,)

comfy/utils.py

@@ -1135,8 +1135,8 @@ def tiled_scale_multidim(samples, function, tile=(64, 64), overlap=8, upscale_am
                 pbar.update(1)
             continue
 
-        out = torch.zeros([s.shape[0], out_channels] + mult_list_upscale(s.shape[2:]), device=output_device)
-        out_div = torch.zeros([s.shape[0], out_channels] + mult_list_upscale(s.shape[2:]), device=output_device)
+        out = output[b:b+1].zero_()
+        out_div = torch.zeros([s.shape[0], 1] + mult_list_upscale(s.shape[2:]), device=output_device)
 
         positions = [range(0, s.shape[d+2] - overlap[d], tile[d] - overlap[d]) if s.shape[d+2] > tile[d] else [0] for d in range(dims)]
 
@@ -1151,7 +1151,7 @@ def tiled_scale_multidim(samples, function, tile=(64, 64), overlap=8, upscale_am
                 upscaled.append(round(get_pos(d, pos)))
 
             ps = function(s_in).to(output_device)
-            mask = torch.ones_like(ps)
+            mask = torch.ones([1, 1] + list(ps.shape[2:]), device=output_device)
 
             for d in range(2, dims + 2):
                 feather = round(get_scale(d - 2, overlap[d - 2]))
@@ -1174,7 +1174,7 @@ def tiled_scale_multidim(samples, function, tile=(64, 64), overlap=8, upscale_am
             if pbar is not None:
                 pbar.update(1)
 
-        output[b:b+1] = out/out_div
+        out.div_(out_div)
     return output
 
 def tiled_scale(samples, function, tile_x=64, tile_y=64, overlap = 8, upscale_amount = 4, out_channels = 3, output_device="cpu", pbar = None):

comfy_api/latest/_io.py

@@ -1353,7 +1353,6 @@ class NodeInfoV1:
     python_module: Any=None
     category: str=None
     output_node: bool=None
-    has_intermediate_output: bool=None
     deprecated: bool=None
     experimental: bool=None
     dev_only: bool=None
@@ -1466,16 +1465,6 @@ class Schema:
 
     Comfy Docs: https://docs.comfy.org/custom-nodes/backend/server_overview#output-node
     """
-    has_intermediate_output: bool=False
-    """Flags this node as having intermediate output that should persist across page refreshes.
-
-    Nodes with this flag behave like output nodes (their UI results are cached and resent
-    to the frontend) but do NOT automatically get added to the execution list. This means
-    they will only execute if they are on the dependency path of a real output node.
-
-    Use this for nodes with interactive/operable UI regions that produce intermediate outputs
-    (e.g., Image Crop, Painter) rather than final outputs (e.g., Save Image).
-    """
     is_deprecated: bool=False
     """Flags a node as deprecated, indicating to users that they should find alternatives to this node."""
     is_experimental: bool=False
@@ -1593,7 +1582,6 @@ class Schema:
             category=self.category,
             description=self.description,
             output_node=self.is_output_node,
-            has_intermediate_output=self.has_intermediate_output,
             deprecated=self.is_deprecated,
             experimental=self.is_experimental,
             dev_only=self.is_dev_only,
@@ -1885,14 +1873,6 @@ class _ComfyNodeBaseInternal(_ComfyNodeInternal):
             cls.GET_SCHEMA()
         return cls._OUTPUT_NODE
 
-    _HAS_INTERMEDIATE_OUTPUT = None
-    @final
-    @classproperty
-    def HAS_INTERMEDIATE_OUTPUT(cls):  # noqa
-        if cls._HAS_INTERMEDIATE_OUTPUT is None:
-            cls.GET_SCHEMA()
-        return cls._HAS_INTERMEDIATE_OUTPUT
-
     _INPUT_IS_LIST = None
     @final
     @classproperty
@@ -1985,8 +1965,6 @@ class _ComfyNodeBaseInternal(_ComfyNodeInternal):
             cls._API_NODE = schema.is_api_node
         if cls._OUTPUT_NODE is None:
             cls._OUTPUT_NODE = schema.is_output_node
-        if cls._HAS_INTERMEDIATE_OUTPUT is None:
-            cls._HAS_INTERMEDIATE_OUTPUT = schema.has_intermediate_output
         if cls._INPUT_IS_LIST is None:
             cls._INPUT_IS_LIST = schema.is_input_list
         if cls._NOT_IDEMPOTENT is None:

comfy_api_nodes/apis/quiver.py

@@ -0,0 +1,43 @@
+from pydantic import BaseModel, Field
+
+
+class QuiverImageObject(BaseModel):
+    url: str = Field(...)
+
+
+class QuiverTextToSVGRequest(BaseModel):
+    model: str = Field(default="arrow-preview")
+    prompt: str = Field(...)
+    instructions: str | None = Field(default=None)
+    references: list[QuiverImageObject] | None = Field(default=None, max_length=4)
+    temperature: float | None = Field(default=None, ge=0, le=2)
+    top_p: float | None = Field(default=None, ge=0, le=1)
+    presence_penalty: float | None = Field(default=None, ge=-2, le=2)
+
+
+class QuiverImageToSVGRequest(BaseModel):
+    model: str = Field(default="arrow-preview")
+    image: QuiverImageObject = Field(...)
+    auto_crop: bool | None = Field(default=None)
+    target_size: int | None = Field(default=None, ge=128, le=4096)
+    temperature: float | None = Field(default=None, ge=0, le=2)
+    top_p: float | None = Field(default=None, ge=0, le=1)
+    presence_penalty: float | None = Field(default=None, ge=-2, le=2)
+
+
+class QuiverSVGResponseItem(BaseModel):
+    svg: str = Field(...)
+    mime_type: str | None = Field(default="image/svg+xml")
+
+
+class QuiverSVGUsage(BaseModel):
+    total_tokens: int | None = Field(default=None)
+    input_tokens: int | None = Field(default=None)
+    output_tokens: int | None = Field(default=None)
+
+
+class QuiverSVGResponse(BaseModel):
+    id: str | None = Field(default=None)
+    created: int | None = Field(default=None)
+    data: list[QuiverSVGResponseItem] = Field(...)
+    usage: QuiverSVGUsage | None = Field(default=None)

comfy_api_nodes/nodes_bytedance.py

@@ -47,6 +47,10 @@ SEEDREAM_MODELS = {
 BYTEPLUS_TASK_ENDPOINT = "/proxy/byteplus/api/v3/contents/generations/tasks"
 BYTEPLUS_TASK_STATUS_ENDPOINT = "/proxy/byteplus/api/v3/contents/generations/tasks"  # + /{task_id}
 
+DEPRECATED_MODELS = {"seedance-1-0-lite-t2v-250428", "seedance-1-0-lite-i2v-250428"}
+
+logger = logging.getLogger(__name__)
+
 
 def get_image_url_from_response(response: ImageTaskCreationResponse) -> str:
     if response.error:
@@ -135,6 +139,7 @@ class ByteDanceImageNode(IO.ComfyNode):
             price_badge=IO.PriceBadge(
                 expr="""{"type":"usd","usd":0.03}""",
             ),
+            is_deprecated=True,
         )
 
     @classmethod
@@ -942,7 +947,7 @@ class ByteDanceImageReferenceNode(IO.ComfyNode):
         ]
         return await process_video_task(
             cls,
-            payload=Image2VideoTaskCreationRequest(model=model, content=x),
+            payload=Image2VideoTaskCreationRequest(model=model, content=x, generate_audio=None),
             estimated_duration=max(1, math.ceil(VIDEO_TASKS_EXECUTION_TIME[model][resolution] * (duration / 10.0))),
         )
 
@@ -952,6 +957,12 @@ async def process_video_task(
     payload: Text2VideoTaskCreationRequest | Image2VideoTaskCreationRequest,
     estimated_duration: int | None,
 ) -> IO.NodeOutput:
+    if payload.model in DEPRECATED_MODELS:
+        logger.warning(
+            "Model '%s' is deprecated and will be deactivated on May 13, 2026. "
+            "Please switch to a newer model. Recommended: seedance-1-0-pro-fast-251015.",
+            payload.model,
+        )
     initial_response = await sync_op(
         cls,
         ApiEndpoint(path=BYTEPLUS_TASK_ENDPOINT, method="POST"),

comfy_api_nodes/nodes_quiver.py

@@ -0,0 +1,291 @@
+from io import BytesIO
+
+from typing_extensions import override
+
+from comfy_api.latest import IO, ComfyExtension
+from comfy_api_nodes.apis.quiver import (
+    QuiverImageObject,
+    QuiverImageToSVGRequest,
+    QuiverSVGResponse,
+    QuiverTextToSVGRequest,
+)
+from comfy_api_nodes.util import (
+    ApiEndpoint,
+    sync_op,
+    upload_image_to_comfyapi,
+    validate_string,
+)
+from comfy_extras.nodes_images import SVG
+
+
+class QuiverTextToSVGNode(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="QuiverTextToSVGNode",
+            display_name="Quiver Text to SVG",
+            category="api node/image/Quiver",
+            description="Generate an SVG from a text prompt using Quiver AI.",
+            inputs=[
+                IO.String.Input(
+                    "prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Text description of the desired SVG output.",
+                ),
+                IO.String.Input(
+                    "instructions",
+                    multiline=True,
+                    default="",
+                    tooltip="Additional style or formatting guidance.",
+                    optional=True,
+                ),
+                IO.Autogrow.Input(
+                    "reference_images",
+                    template=IO.Autogrow.TemplatePrefix(
+                        IO.Image.Input("image"),
+                        prefix="ref_",
+                        min=0,
+                        max=4,
+                    ),
+                    tooltip="Up to 4 reference images to guide the generation.",
+                    optional=True,
+                ),
+                IO.DynamicCombo.Input(
+                    "model",
+                    options=[
+                        IO.DynamicCombo.Option(
+                            "arrow-preview",
+                            [
+                                IO.Float.Input(
+                                    "temperature",
+                                    default=1.0,
+                                    min=0.0,
+                                    max=2.0,
+                                    step=0.1,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Randomness control. Higher values increase randomness.",
+                                    advanced=True,
+                                ),
+                                IO.Float.Input(
+                                    "top_p",
+                                    default=1.0,
+                                    min=0.05,
+                                    max=1.0,
+                                    step=0.05,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Nucleus sampling parameter.",
+                                    advanced=True,
+                                ),
+                                IO.Float.Input(
+                                    "presence_penalty",
+                                    default=0.0,
+                                    min=-2.0,
+                                    max=2.0,
+                                    step=0.1,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Token presence penalty.",
+                                    advanced=True,
+                                ),
+                            ],
+                        ),
+                    ],
+                    tooltip="Model to use for SVG generation.",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    control_after_generate=True,
+                    tooltip="Seed to determine if node should re-run; "
+                    "actual results are nondeterministic regardless of seed.",
+                ),
+            ],
+            outputs=[
+                IO.SVG.Output(),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                expr="""{"type":"usd","usd":0.429}""",
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        prompt: str,
+        model: dict,
+        seed: int,
+        instructions: str = None,
+        reference_images: IO.Autogrow.Type = None,
+    ) -> IO.NodeOutput:
+        validate_string(prompt, strip_whitespace=False, min_length=1)
+
+        references = None
+        if reference_images:
+            references = []
+            for key in reference_images:
+                url = await upload_image_to_comfyapi(cls, reference_images[key])
+                references.append(QuiverImageObject(url=url))
+            if len(references) > 4:
+                raise ValueError("Maximum 4 reference images are allowed.")
+
+        instructions_val = instructions.strip() if instructions else None
+        if instructions_val == "":
+            instructions_val = None
+
+        response = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/quiver/v1/svgs/generations", method="POST"),
+            response_model=QuiverSVGResponse,
+            data=QuiverTextToSVGRequest(
+                model=model["model"],
+                prompt=prompt,
+                instructions=instructions_val,
+                references=references,
+                temperature=model.get("temperature"),
+                top_p=model.get("top_p"),
+                presence_penalty=model.get("presence_penalty"),
+            ),
+        )
+
+        svg_data = [BytesIO(item.svg.encode("utf-8")) for item in response.data]
+        return IO.NodeOutput(SVG(svg_data))
+
+
+class QuiverImageToSVGNode(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="QuiverImageToSVGNode",
+            display_name="Quiver Image to SVG",
+            category="api node/image/Quiver",
+            description="Vectorize a raster image into SVG using Quiver AI.",
+            inputs=[
+                IO.Image.Input(
+                    "image",
+                    tooltip="Input image to vectorize.",
+                ),
+                IO.Boolean.Input(
+                    "auto_crop",
+                    default=False,
+                    tooltip="Automatically crop to the dominant subject.",
+                ),
+                IO.DynamicCombo.Input(
+                    "model",
+                    options=[
+                        IO.DynamicCombo.Option(
+                            "arrow-preview",
+                            [
+                                IO.Int.Input(
+                                    "target_size",
+                                    default=1024,
+                                    min=128,
+                                    max=4096,
+                                    tooltip="Square resize target in pixels.",
+                                ),
+                                IO.Float.Input(
+                                    "temperature",
+                                    default=1.0,
+                                    min=0.0,
+                                    max=2.0,
+                                    step=0.1,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Randomness control. Higher values increase randomness.",
+                                    advanced=True,
+                                ),
+                                IO.Float.Input(
+                                    "top_p",
+                                    default=1.0,
+                                    min=0.05,
+                                    max=1.0,
+                                    step=0.05,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Nucleus sampling parameter.",
+                                    advanced=True,
+                                ),
+                                IO.Float.Input(
+                                    "presence_penalty",
+                                    default=0.0,
+                                    min=-2.0,
+                                    max=2.0,
+                                    step=0.1,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Token presence penalty.",
+                                    advanced=True,
+                                ),
+                            ],
+                        ),
+                    ],
+                    tooltip="Model to use for SVG vectorization.",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    control_after_generate=True,
+                    tooltip="Seed to determine if node should re-run; "
+                    "actual results are nondeterministic regardless of seed.",
+                ),
+            ],
+            outputs=[
+                IO.SVG.Output(),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                expr="""{"type":"usd","usd":0.429}""",
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        image,
+        auto_crop: bool,
+        model: dict,
+        seed: int,
+    ) -> IO.NodeOutput:
+        image_url = await upload_image_to_comfyapi(cls, image)
+
+        response = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/quiver/v1/svgs/vectorizations", method="POST"),
+            response_model=QuiverSVGResponse,
+            data=QuiverImageToSVGRequest(
+                model=model["model"],
+                image=QuiverImageObject(url=image_url),
+                auto_crop=auto_crop if auto_crop else None,
+                target_size=model.get("target_size"),
+                temperature=model.get("temperature"),
+                top_p=model.get("top_p"),
+                presence_penalty=model.get("presence_penalty"),
+            ),
+        )
+
+        svg_data = [BytesIO(item.svg.encode("utf-8")) for item in response.data]
+        return IO.NodeOutput(SVG(svg_data))
+
+
+class QuiverExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[IO.ComfyNode]]:
+        return [
+            QuiverTextToSVGNode,
+            QuiverImageToSVGNode,
+        ]
+
+
+async def comfy_entrypoint() -> QuiverExtension:
+    return QuiverExtension()

comfy_execution/graph.py

@@ -118,13 +118,6 @@ class TopologicalSort:
         class_def = nodes.NODE_CLASS_MAPPINGS[class_type]
         return get_input_info(class_def, input_name)
 
-    def is_intermediate_output(self, node_id):
-        class_type = self.dynprompt.get_node(node_id)["class_type"]
-        class_def = nodes.NODE_CLASS_MAPPINGS.get(class_type)
-        if class_def is None:
-            return False
-        return hasattr(class_def, 'HAS_INTERMEDIATE_OUTPUT') and class_def.HAS_INTERMEDIATE_OUTPUT == True
-
     def make_input_strong_link(self, to_node_id, to_input):
         inputs = self.dynprompt.get_node(to_node_id)["inputs"]
         if to_input not in inputs:
@@ -136,7 +129,7 @@ class TopologicalSort:
         self.add_strong_link(from_node_id, from_socket, to_node_id)
 
     def add_strong_link(self, from_node_id, from_socket, to_node_id):
-        if not self.is_cached(from_node_id) or self.is_intermediate_output(from_node_id):
+        if not self.is_cached(from_node_id):
             self.add_node(from_node_id)
             if to_node_id not in self.blocking[from_node_id]:
                 self.blocking[from_node_id][to_node_id] = {}
@@ -166,7 +159,7 @@ class TopologicalSort:
                     _, _, input_info = self.get_input_info(unique_id, input_name)
                     is_lazy = input_info is not None and "lazy" in input_info and input_info["lazy"]
                     if (include_lazy or not is_lazy):
-                        if not self.is_cached(from_node_id) or self.is_intermediate_output(from_node_id):
+                        if not self.is_cached(from_node_id):
                             node_ids.append(from_node_id)
                         links.append((from_node_id, from_socket, unique_id))
 
@@ -284,8 +277,6 @@ class ExecutionList(TopologicalSort):
             class_def = nodes.NODE_CLASS_MAPPINGS[class_type]
             if hasattr(class_def, 'OUTPUT_NODE') and class_def.OUTPUT_NODE == True:
                 return True
-            if hasattr(class_def, 'HAS_INTERMEDIATE_OUTPUT') and class_def.HAS_INTERMEDIATE_OUTPUT == True:
-                return True
             return False
 
         # If an available node is async, do that first.

comfy_extras/nodes_canny.py

@@ -3,6 +3,7 @@ from typing_extensions import override
 
 import comfy.model_management
 from comfy_api.latest import ComfyExtension, io
+import torch
 
 
 class Canny(io.ComfyNode):
@@ -29,8 +30,8 @@ class Canny(io.ComfyNode):
 
     @classmethod
     def execute(cls, image, low_threshold, high_threshold) -> io.NodeOutput:
-        output = canny(image.to(comfy.model_management.get_torch_device()).movedim(-1, 1), low_threshold, high_threshold)
-        img_out = output[1].to(comfy.model_management.intermediate_device()).repeat(1, 3, 1, 1).movedim(1, -1)
+        output = canny(image.to(device=comfy.model_management.get_torch_device(), dtype=torch.float32).movedim(-1, 1), low_threshold, high_threshold)
+        img_out = output[1].to(device=comfy.model_management.intermediate_device(), dtype=comfy.model_management.intermediate_dtype()).repeat(1, 3, 1, 1).movedim(1, -1)
         return io.NodeOutput(img_out)
 
 

comfy_extras/nodes_context_windows.py

@@ -27,8 +27,8 @@ class ContextWindowsManualNode(io.ComfyNode):
                 io.Combo.Input("fuse_method", options=comfy.context_windows.ContextFuseMethods.LIST_STATIC, default=comfy.context_windows.ContextFuseMethods.PYRAMID, tooltip="The method to use to fuse the context windows."),
                 io.Int.Input("dim", min=0, max=5, default=0, tooltip="The dimension to apply the context windows to."),
                 io.Boolean.Input("freenoise", default=False, tooltip="Whether to apply FreeNoise noise shuffling, improves window blending."),
-                #io.String.Input("cond_retain_index_list", default="", tooltip="List of latent indices to retain in the conditioning tensors for each window, for example setting this to '0' will use the initial start image for each window."),
-                #io.Boolean.Input("split_conds_to_windows", default=False, tooltip="Whether to split multiple conditionings (created by ConditionCombine) to each window based on region index."),
+                io.String.Input("cond_retain_index_list", default="", tooltip="List of latent indices to retain in the conditioning tensors for each window, for example setting this to '0' will use the initial start image for each window."),
+                io.Boolean.Input("split_conds_to_windows", default=False, tooltip="Whether to split multiple conditionings (created by ConditionCombine) to each window based on region index."),
             ],
             outputs=[
                 io.Model.Output(tooltip="The model with context windows applied during sampling."),

comfy_extras/nodes_glsl.py

@@ -762,7 +762,6 @@ class GLSLShader(io.ComfyNode):
                 "Apply GLSL ES fragment shaders to images. "
                 "u_resolution (vec2) is always available."
             ),
-            has_intermediate_output=True,
             inputs=[
                 io.String.Input(
                     "fragment_shader",

comfy_extras/nodes_images.py

@@ -59,7 +59,6 @@ class ImageCropV2(IO.ComfyNode):
             display_name="Image Crop",
             category="image/transform",
             essentials_category="Image Tools",
-            has_intermediate_output=True,
             inputs=[
                 IO.Image.Input("image"),
                 IO.BoundingBox.Input("crop_region", component="ImageCrop"),

comfy_extras/nodes_painter.py

@@ -30,7 +30,6 @@ class PainterNode(io.ComfyNode):
             node_id="Painter",
             display_name="Painter",
             category="image",
-            has_intermediate_output=True,
             inputs=[
                 io.Image.Input(
                     "image",

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.17.0"
+__version__ = "0.18.0"

nodes.py

@@ -1966,9 +1966,11 @@ class EmptyImage:
     CATEGORY = "image"
 
     def generate(self, width, height, batch_size=1, color=0):
-        r = torch.full([batch_size, height, width, 1], ((color >> 16) & 0xFF) / 0xFF)
-        g = torch.full([batch_size, height, width, 1], ((color >> 8) & 0xFF) / 0xFF)
-        b = torch.full([batch_size, height, width, 1], ((color) & 0xFF) / 0xFF)
+        dtype = comfy.model_management.intermediate_dtype()
+        device = comfy.model_management.intermediate_device()
+        r = torch.full([batch_size, height, width, 1], ((color >> 16) & 0xFF) / 0xFF, device=device, dtype=dtype)
+        g = torch.full([batch_size, height, width, 1], ((color >> 8) & 0xFF) / 0xFF, device=device, dtype=dtype)
+        b = torch.full([batch_size, height, width, 1], ((color) & 0xFF) / 0xFF, device=device, dtype=dtype)
         return (torch.cat((r, g, b), dim=-1), )
 
 class ImagePadForOutpaint:

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.17.0"
+version = "0.18.0"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.10"

requirements.txt

@@ -1,4 +1,4 @@
-comfyui-frontend-package==1.41.20
+comfyui-frontend-package==1.41.21
 comfyui-workflow-templates==0.9.26
 comfyui-embedded-docs==0.4.3
 torch

server.py

@@ -709,9 +709,6 @@ class PromptServer():
             else:
                 info['output_node'] = False
 
-            if hasattr(obj_class, 'HAS_INTERMEDIATE_OUTPUT') and obj_class.HAS_INTERMEDIATE_OUTPUT == True:
-                info['has_intermediate_output'] = True
-
             if hasattr(obj_class, 'CATEGORY'):
                 info['category'] = obj_class.CATEGORY