comfyanonymous/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-02-27 02:14:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'master' into dr-support-pip-cm

Browse Source
This commit is contained in:
Dr.Lt.Data
2025-08-21 06:44:35 +09:00
parent 4b1aac74bb e73a9dbe30
commit d7b4f45c5b
8 changed files with 17 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
comfy/ldm/qwen_image/model.py
View File

@@ -347,10 +347,10 @@ class QwenImageTransformer2DModel(nn.Module):
h_offset = ((h_offset + (patch_size // 2)) // patch_size)
w_offset = ((w_offset + (patch_size // 2)) // patch_size)
img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype)
img_ids = torch.zeros((h_len, w_len, 3), device=x.device)
img_ids[:, :, 0] = img_ids[:, :, 1] + index
img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1)
img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0)
img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1) - (h_len // 2)
img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0) - (w_len // 2)
return hidden_states, repeat(img_ids, "h w c -> b (h w) c", b=bs), orig_shape
def forward(
@@ -396,10 +396,11 @@ class QwenImageTransformer2DModel(nn.Module):
hidden_states = torch.cat([hidden_states, kontext], dim=1)
img_ids = torch.cat([img_ids, kontext_ids], dim=1)
txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size), ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size)))
txt_ids = torch.linspace(txt_start, txt_start + context.shape[1], steps=context.shape[1], device=x.device, dtype=x.dtype).reshape(1, -1, 1).repeat(x.shape[0], 1, 3)
txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size) // 2, ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size) // 2))
txt_ids = torch.arange(txt_start, txt_start + context.shape[1], device=x.device).reshape(1, -1, 1).repeat(x.shape[0], 1, 3)
ids = torch.cat((txt_ids, img_ids), dim=1)
image_rotary_emb = self.pe_embedder(ids).squeeze(1).unsqueeze(2).to(x.dtype)
del ids, txt_ids, img_ids
hidden_states = self.img_in(hidden_states)
encoder_hidden_states = self.txt_norm(encoder_hidden_states)

5
comfy/sd1_clip.py
View File

@@ -534,7 +534,10 @@ class SDTokenizer:
min_padding = tokenizer_options.get("{}_min_padding".format(self.embedding_key), self.min_padding)
text = escape_important(text)
parsed_weights = token_weights(text, 1.0)
if kwargs.get("disable_weights", False):
parsed_weights = [(text, 1.0)]
else:
parsed_weights = token_weights(text, 1.0)
# tokenize words
tokens = []

4
comfy/text_encoders/qwen_image.py
View File

@@ -15,7 +15,7 @@ class QwenImageTokenizer(sd1_clip.SD1Tokenizer):
def __init__(self, embedding_directory=None, tokenizer_data={}):
super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="qwen25_7b", tokenizer=Qwen25_7BVLITokenizer)
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"
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):
if llama_template is None:
@@ -25,7 +25,7 @@ class QwenImageTokenizer(sd1_clip.SD1Tokenizer):
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, **kwargs)
tokens = super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, disable_weights=True, **kwargs)
key_name = next(iter(tokens))
embed_count = 0
qwen_tokens = tokens[key_name]