A workflow uses the latest Qwen Image Edit F2P lora from DiffSynth-Studio to swap faces, while maintaining an extremely high fidelity. Support expression transfer.
Try the F2P model online for free: https://www.runninghub.ai/post/1985156515172667394
Please check the video guide before you start using this workflow:
If you think the video is helpful, please like, share the video, and subscribe for more. I will update with more interesting things. Thank you.
Description
Update the model to Qwen Image Edit 2511, add expression transfer and manual masking for precise face swap.
FAQ
Comments (12)
this workflow is great. thanks. is there any solution for an image with multiple faces that need swapped?
does not work. i tried using the lower option where i highlight the mask. i did not change anything else, just added the photos and that's it. absolutely does not work. i couldn't get the workflow to work. it's junk. it doesn't work.
This worked great, QwenVL was giving me a lot of errors so i had to bypass the whole expression transfer group, but even without it did a fantastic job, thanks for this workflow and others, you always have some cool workflows i don't see anywhere else.
For me the QwenVL node is halting the workflow with no message (just red outline). What should I check to try to fix it?
attention_mode=auto
preset_prompt=Detailed Description
max_tokens=512
incredible. Thanks !
@sdktertiaire2 does the qwenvl node work for you. The workflow stops unless I bypass it
Thanks you this actually works (had to bypass the QwenVL)
I had various issues in Human-parts-ultra and QwenVL so debugging the workflow was necessary. Especially, you don't need to run the ultra node in gpu, which created an access violation for me, but the node runs gpu anyway even with cpu selected. It took a while to fix the bug. But in the end, I could see the excellent result of expression transfer that I've seen rarely. Thanks for sharing.
can you share how you fixed it please? getting all kinds of errors
@zdakd829473506 Well, as it's been a week, I barely remember anything but I will try. As mentioned above, I didn't have many errors but one complaining cuda error in human-parts-ultra node. Because setting cpu for the node still gave the same error, I edited the python file for the node: "...\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui_layerstyle\py\human_parts_ultra.py" and I replaced the following comment-out line with uncommented one so it would not run gpu for that node:
# model = ort.InferenceSession(model_path, providers=['TensorrtExecutionProvider', 'CUDAExecutionProvider', 'CPUExecutionProvider'])
model = ort.InferenceSession(model_path, providers=[ 'CPUExecutionProvider'])
About QwenVL, I had an issue with running non-native models but you would not have this issue if you are using native qwen models.
How to use the manual masking version?