Add warning for users who disable dynamic vram.

There was an issue where the resample split was too early and dropped one
of the rolling convolutions a frame early. This is most noticable as a
lighting/color change between pixel frames 5->6 (latent 2->3), or as a
lighting change between the first and last frame in an FLF wan flow.

comfy/ldm/flux/model.py

@@ -386,7 +386,7 @@ class Flux(nn.Module):
                     h = max(h, ref.shape[-2] + h_offset)
                     w = max(w, ref.shape[-1] + w_offset)
 
-                kontext, kontext_ids = self.process_img(ref, index=index, h_offset=h_offset, w_offset=w_offset)
+                kontext, kontext_ids = self.process_img(ref, index=index, h_offset=h_offset, w_offset=w_offset, transformer_options=transformer_options)
                 img = torch.cat([img, kontext], dim=1)
                 img_ids = torch.cat([img_ids, kontext_ids], dim=1)
                 ref_num_tokens.append(kontext.shape[1])

comfy/ldm/wan/vae.py

@@ -376,11 +376,16 @@ class Decoder3d(nn.Module):
             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]):
+        if feat_cache is not None:
+            x = layer(x, feat_cache, feat_idx)
+        else:
+            x = layer(x)
+
+        if isinstance(layer, Resample) and layer.mode == 'upsample3d' and x.shape[2] > 2:
+            for frame_idx in range(0, x.shape[2], 2):
                 self.run_up(
-                    layer_idx,
-                    [x[:, :, frame_idx:frame_idx + 1, :, :]],
+                    layer_idx + 1,
+                    [x[:, :, frame_idx:frame_idx + 2, :, :]],
                     feat_cache,
                     feat_idx.copy(),
                     out_chunks,
@@ -388,11 +393,6 @@ class Decoder3d(nn.Module):
             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)

comfy/model_base.py

@@ -937,9 +937,10 @@ class LongCatImage(Flux):
         transformer_options = transformer_options.copy()
         rope_opts = transformer_options.get("rope_options", {})
         rope_opts = dict(rope_opts)
+        pe_len = float(c_crossattn.shape[1]) if c_crossattn is not None else 512.0
         rope_opts.setdefault("shift_t", 1.0)
-        rope_opts.setdefault("shift_y", 512.0)
-        rope_opts.setdefault("shift_x", 512.0)
+        rope_opts.setdefault("shift_y", pe_len)
+        rope_opts.setdefault("shift_x", pe_len)
         transformer_options["rope_options"] = rope_opts
         return super()._apply_model(x, t, c_concat, c_crossattn, control, transformer_options, **kwargs)
 

comfy/sample.py

@@ -8,12 +8,12 @@ import comfy.nested_tensor
 
 def prepare_noise_inner(latent_image, generator, noise_inds=None):
     if noise_inds is None:
-        return torch.randn(latent_image.size(), dtype=latent_image.dtype, layout=latent_image.layout, generator=generator, device="cpu")
+        return torch.randn(latent_image.size(), dtype=torch.float32, layout=latent_image.layout, generator=generator, device="cpu").to(dtype=latent_image.dtype)
 
     unique_inds, inverse = np.unique(noise_inds, return_inverse=True)
     noises = []
     for i in range(unique_inds[-1]+1):
-        noise = torch.randn([1] + list(latent_image.size())[1:], dtype=latent_image.dtype, layout=latent_image.layout, generator=generator, device="cpu")
+        noise = torch.randn([1] + list(latent_image.size())[1:], dtype=torch.float32, layout=latent_image.layout, generator=generator, device="cpu").to(dtype=latent_image.dtype)
         if i in unique_inds:
             noises.append(noise)
     noises = [noises[i] for i in inverse]

comfy/samplers.py

@@ -985,8 +985,8 @@ class CFGGuider:
         self.inner_model, self.conds, self.loaded_models = comfy.sampler_helpers.prepare_sampling(self.model_patcher, noise.shape, self.conds, self.model_options)
         device = self.model_patcher.load_device
 
-        noise = noise.to(device)
-        latent_image = latent_image.to(device)
+        noise = noise.to(device=device, dtype=torch.float32)
+        latent_image = latent_image.to(device=device, dtype=torch.float32)
         sigmas = sigmas.to(device)
         cast_to_load_options(self.model_options, device=device, dtype=self.model_patcher.model_dtype())
 
@@ -1028,6 +1028,7 @@ class CFGGuider:
                 denoise_mask, _ = comfy.utils.pack_latents(denoise_masks)
             else:
                 denoise_mask = denoise_masks[0]
+            denoise_mask = denoise_mask.float()
 
         self.conds = {}
         for k in self.original_conds:

comfy/text_encoders/llama.py

@@ -1028,12 +1028,19 @@ class Qwen25_7BVLI(BaseLlama, BaseGenerate, torch.nn.Module):
                 grid = e.get("extra", None)
                 start = e.get("index")
                 if position_ids is None:
-                    position_ids = torch.zeros((3, embeds.shape[1]), device=embeds.device)
+                    position_ids = torch.ones((3, embeds.shape[1]), device=embeds.device, dtype=torch.long)
                     position_ids[:, :start] = torch.arange(0, start, device=embeds.device)
                 end = e.get("size") + start
                 len_max = int(grid.max()) // 2
                 start_next = len_max + start
-                position_ids[:, end:] = torch.arange(start_next + offset, start_next + (embeds.shape[1] - end) + offset, device=embeds.device)
+                if attention_mask is not None:
+                    # Assign compact sequential positions to attended tokens only,
+                    # skipping over padding so post-padding tokens aren't inflated.
+                    after_mask = attention_mask[0, end:]
+                    text_positions = after_mask.cumsum(0) - 1 + start_next + offset
+                    position_ids[:, end:] = torch.where(after_mask.bool(), text_positions, position_ids[0, end:])
+                else:
+                    position_ids[:, end:] = torch.arange(start_next + offset, start_next + (embeds.shape[1] - end) + offset, device=embeds.device)
                 position_ids[0, start:end] = start + offset
                 max_d = int(grid[0][1]) // 2
                 position_ids[1, start:end] = torch.arange(start + offset, start + max_d + offset, device=embeds.device).unsqueeze(1).repeat(1, math.ceil((end - start) / max_d)).flatten(0)[:end - start]

comfy/text_encoders/longcat_image.py

@@ -64,7 +64,13 @@ class LongCatImageBaseTokenizer(Qwen25_7BVLITokenizer):
         return [output]
 
 
+IMAGE_PAD_TOKEN_ID = 151655
+
 class LongCatImageTokenizer(sd1_clip.SD1Tokenizer):
+    T2I_PREFIX = "<|im_start|>system\nAs an image captioning expert, generate a descriptive text prompt based on an image content, suitable for input to a text-to-image model.<|im_end|>\n<|im_start|>user\n"
+    EDIT_PREFIX = "<|im_start|>system\nAs an image editing expert, first analyze the content and attributes of the input image(s). Then, based on the user's editing instructions, clearly and precisely determine how to modify the given image(s), ensuring that only the specified parts are altered and all other aspects remain consistent with the original(s).<|im_end|>\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>"
+    SUFFIX = "<|im_end|>\n<|im_start|>assistant\n"
+
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         super().__init__(
             embedding_directory=embedding_directory,
@@ -72,10 +78,8 @@ class LongCatImageTokenizer(sd1_clip.SD1Tokenizer):
             name="qwen25_7b",
             tokenizer=LongCatImageBaseTokenizer,
         )
-        self.longcat_template_prefix = "<|im_start|>system\nAs an image captioning expert, generate a descriptive text prompt based on an image content, suitable for input to a text-to-image model.<|im_end|>\n<|im_start|>user\n"
-        self.longcat_template_suffix = "<|im_end|>\n<|im_start|>assistant\n"
 
-    def tokenize_with_weights(self, text, return_word_ids=False, **kwargs):
+    def tokenize_with_weights(self, text, return_word_ids=False, images=None, **kwargs):
         skip_template = False
         if text.startswith("<|im_start|>"):
             skip_template = True
@@ -90,11 +94,14 @@ class LongCatImageTokenizer(sd1_clip.SD1Tokenizer):
                 text, return_word_ids=return_word_ids, disable_weights=True, **kwargs
             )
         else:
+            has_images = images is not None and len(images) > 0
+            template_prefix = self.EDIT_PREFIX if has_images else self.T2I_PREFIX
+
             prefix_ids = base_tok.tokenizer(
-                self.longcat_template_prefix, add_special_tokens=False
+                template_prefix, add_special_tokens=False
             )["input_ids"]
             suffix_ids = base_tok.tokenizer(
-                self.longcat_template_suffix, add_special_tokens=False
+                self.SUFFIX, add_special_tokens=False
             )["input_ids"]
 
             prompt_tokens = base_tok.tokenize_with_weights(
@@ -106,6 +113,14 @@ class LongCatImageTokenizer(sd1_clip.SD1Tokenizer):
             suffix_pairs = [(t, 1.0) for t in suffix_ids]
 
             combined = prefix_pairs + prompt_pairs + suffix_pairs
+
+            if has_images:
+                embed_count = 0
+                for i in range(len(combined)):
+                    if combined[i][0] == IMAGE_PAD_TOKEN_ID and embed_count < len(images):
+                        combined[i] = ({"type": "image", "data": images[embed_count], "original_type": "image"}, combined[i][1])
+                        embed_count += 1
+
             tokens = {"qwen25_7b": [combined]}
 
         return tokens

comfy/text_encoders/qwen_vl.py

@@ -425,4 +425,7 @@ class Qwen2VLVisionTransformer(nn.Module):
             hidden_states = block(hidden_states, position_embeddings, cu_seqlens_now, optimized_attention=optimized_attention)
 
         hidden_states = self.merger(hidden_states)
+        # Potentially important for spatially precise edits. This is present in the HF implementation.
+        reverse_indices = torch.argsort(window_index)
+        hidden_states = hidden_states[reverse_indices, :]
         return hidden_states

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)
 
 

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.18.0"
+__version__ = "0.18.1"

main.py

@@ -471,6 +471,9 @@ if __name__ == "__main__":
     if sys.version_info.major == 3 and sys.version_info.minor < 10:
         logging.warning("WARNING: You are using a python version older than 3.10, please upgrade to a newer one. 3.12 and above is recommended.")
 
+    if args.disable_dynamic_vram:
+        logging.warning("Dynamic vram disabled with argument. If you have any issues with dynamic vram enabled please give us a detailed reports as this argument will be removed soon.")
+
     event_loop, _, start_all_func = start_comfyui()
     try:
         x = start_all_func()

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.18.0"
+version = "0.18.1"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.10"