Qwen-Image-Edit-2511作为一款性能出色的图像编辑模型，在ComfyUI中部署时却受限于显存资源。本文针对4090显卡（24G显存）场景，分享量化模型的部署流程、关键避坑点，以及不同采样步数下的效果对比，帮助大家快速落地实践。
一、前置准备：ComfyUI安装
ComfyUI基础安装流程此处不赘述，推荐参考官方中文指南，步骤清晰且适配Linux环境：ComfyUI Linux安装官方指南
二、核心问题：显存限制与解决方案
4090显卡的24G显存无法承载Qwen-Image-Edit-2511原始模型（显存溢出），因此必须使用量化模型。考虑到外网访问和下载限制，本文提供基于hugging-face镜像 和modelscope的国内可访问下载链接，同时明确各模型的存放路径（直接关系到模型能否正常加载）。
2.1 量化模型下载清单（含路径+命令）
所有模型需下载至ComfyUI对应目录，以下是完整的路径说明和wget下载命令（复制到终端直接执行即可）：
1. LoRA模型（路径：ComfyUI/models/loras）
bash
wget https://hf-mirror.com/lightx2v/Qwen-Image-Edit-2511-Lightning/resolve/main/Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors2. VAE模型（路径：ComfyUI/models/vae）
bash
wget https://hf-mirror.com/Comfy-Org/Qwen-Image_ComfyUI/resolve/main/split_files/vae/qwen_image_vae.safetensors3. UNet模型（路径：ComfyUI/models/unet）
bash
wget "https://modelscope.cn/api/v1/models/unsloth/Qwen-Image-Edit-2511-GGUF/repo?Revision=master&FilePath=qwen-image-edit-2511-Q4_K_M.gguf" -O qwen-image-edit-2511-Q4_K_M.gguf4. CLIP模型（路径：ComfyUI/models/clip）
bash
# 主模型文件
wget -c "https://modelscope.cn/api/v1/models/unsloth/Qwen2.5-VL-7B-Instruct-GGUF/repo?Revision=master&FilePath=Qwen2.5-VL-7B-Instruct-Q4_K_M.gguf" -O Qwen2.5-VL-7B-Instruct-Q4_K_M.gguf
# 关键依赖文件（必下！）
wget -c "https://modelscope.cn/api/v1/models/unsloth/Qwen2.5-VL-7B-Instruct-GGUF/repo?Revision=master&FilePath=mmproj-F16.gguf" -O Qwen2.5-VL-7B-Instruct-mmproj-BF16.gguf2.2 致命坑点：缺失mmproj文件导致的报错解决方案
⚠️ 重点提醒：CLIP模型对应的mmproj文件是必下载项！缺失该文件会直接导致图像编辑时出现「矩阵维度不匹配」的致命错误，排查过程耗时耗力。
我首次部署时因遗漏该文件，出现如下报错（核心信息：mat1 and mat2 shapes cannot be multiplied）：
Prompt executed in 10.11 seconds
got prompt
!!! Exception during processing !!! mat1 and mat2 shapes cannot be multiplied (748x1280 and 3840x1280)
Traceback (most recent call last):
File "/root/comfy/ComfyUI/execution.py", line 516, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy/ComfyUI/execution.py", line 330, in get_output_data
return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy/ComfyUI/execution.py", line 304, in _async_map_node_over_list
await process_inputs(input_dict, i)
File "/root/comfy/ComfyUI/execution.py", line 292, in process_inputs
result = f(**inputs)
^^^^^^^^^^^
File "/root/comfy/ComfyUI/comfy_api/internal/__init__.py", line 149, in wrapped_func
return method(locked_class, **inputs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy/ComfyUI/comfy_api/latest/_io.py", line 1520, in EXECUTE_NORMALIZED
to_return = cls.execute(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy/ComfyUI/comfy_extras/nodes_qwen.py", line 103, in execute
conditioning = clip.encode_from_tokens_scheduled(tokens)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy/ComfyUI/comfy/sd.py", line 207, in encode_from_tokens_scheduled
pooled_dict = self.encode_from_tokens(tokens, return_pooled=return_pooled, return_dict=True)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy/ComfyUI/comfy/sd.py", line 271, in encode_from_tokens
o = self.cond_stage_model.encode_token_weights(tokens)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy/ComfyUI/comfy/text_encoders/qwen_image.py", line 62, in encode_token_weights
out, pooled, extra = super().encode_token_weights(token_weight_pairs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy/ComfyUI/comfy/sd1_clip.py", line 704, in encode_token_weights
out = getattr(self, self.clip).encode_token_weights(token_weight_pairs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy/ComfyUI/comfy/sd1_clip.py", line 45, in encode_token_weights
o = self.encode(to_encode)
^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy/ComfyUI/comfy/sd1_clip.py", line 297, in encode
return self(tokens)
^^^^^^^^^^^^
File "/root/comfy-env/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy-env/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy/ComfyUI/comfy/sd1_clip.py", line 257, in forward
embeds, attention_mask, num_tokens, embeds_info = self.process_tokens(tokens, device)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy/ComfyUI/comfy/sd1_clip.py", line 219, in process_tokens
emb, extra = self.transformer.preprocess_embed(emb, device=device)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy/ComfyUI/comfy/text_encoders/llama.py", line 593, in preprocess_embed
return self.visual(image.to(device, dtype=torch.float32), grid), grid
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy-env/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy-env/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy/ComfyUI/comfy/text_encoders/qwen_vl.py", line 425, in forward
hidden_states = block(hidden_states, position_embeddings, cu_seqlens_now, optimized_attention=optimized_attention)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy-env/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy-env/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy/ComfyUI/comfy/text_encoders/qwen_vl.py", line 252, in forward
hidden_states = self.attn(hidden_states, position_embeddings, cu_seqlens, optimized_attention)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy-env/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy-env/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy/ComfyUI/comfy/text_encoders/qwen_vl.py", line 195, in forward
qkv = self.qkv(hidden_states)
^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy-env/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy-env/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy/ComfyUI/comfy/ops.py", line 164, in forward
return self.forward_comfy_cast_weights(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy/ComfyUI/custom_nodes/ComfyUI-GGUF/ops.py", line 217, in forward_comfy_cast_weights
out = super().forward_comfy_cast_weights(input, *args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/comfy/ComfyUI/comfy/ops.py", line 157, in forward_comfy_cast_weights
x = torch.nn.functional.linear(input, weight, bias)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: mat1 and mat2 shapes cannot be multiplied (748x1280 and 3840x1280)最终通过GitHubissue找到解决方案（感谢开源社区）：TextEncodeQwenImageEdit mat1 and mat2 shapes cannot be multiplied 问题解决方案，核心就是补全mmproj文件。建议大家直接按上文清单下载，避免重复踩坑。
三、工作流配置与效果测试
模型部署完成后，需配置对应的工作流。以下是我测试用的工作流截图（可直接参考复刻）：
本次测试以「三图编辑」为场景，重点验证不同K采样器步数对输出效果的影响，测试环境为4090显卡+Linux系统，具体结果如下：
3.1 20步采样：速度快但效果差
- 运行时长：1分40秒
- 效果问题：人物手臂存在明显割裂感；人物面部失真严重（如"马爸爸"面部完全识别不出）
- 效果截图：
3.2 40步采样：效果略有提升但仍有缺陷
- 运行时长：4分37秒
- 效果问题：手部与手臂的割裂感未完全解决，仍存在明显衔接瑕疵
- 效果截图：
3.3 60步采样：效果达标但耗时增加
- 运行时长：6分57秒
- 效果表现：手臂衔接问题基本解决；但人物面部仍与原角色有较大差异，且出现非预期的衣物颜色变化（如浅灰色衣物变为黑色）
- 效果截图：
四、总结与后续优化方向
- 4090显卡运行Qwen-Image-Edit-2511需优先选择量化模型，按本文提供的国内镜像链接下载可规避网络问题，且mmproj文件不可遗漏；
- 采样步数与效果、速度呈正相关：20步适合快速预览，60步可解决核心瑕疵，但需接受更长耗时和人物脸部变更；
- 后续优化方向：可尝试调整工作流中的提示词精度、优化流程参数，或测试更高精度的量化模型（如Q2_K等），平衡效果与耗时。
如果大家在部署过程中遇到其他问题，欢迎在评论区交流～