search

wenquanlu / HandRefiner

MIT License
701 stars 30 forks source link

unstable refined results #36

Open Aoiryo opened 3 days ago

Aoiryo commented 3 days ago

Hi, thanks for your amazing works! I am trying the model on ComfyUI, and I got the following results from different seeds. The results seem not as good as those from the paper, what could be potential problems? Am I missing something or do you have any guesses on why this is happening?

Any suggestions would be helpful, thanks! 截屏2024-06-27 03 07 30

Aoiryo commented 3 days ago

BTW, I am using: materpiece, good quality, photorealistic, a Chinese woman waving her hands toward the camera as my positive prompt, and worst quality, bad quality, nsfw, CG, as the negative one

wenquanlu commented 3 days ago

Thanks for your enquiry. Yeah looks like definitely something is wrong, what weight are you using, and what is the control strength? (it would be helpful if you can screenshot your entire workflow so I can try to debug)

I personally have never duplicated this project in ComfyUI, so if you try to duplicate it from scratch, it is very likely you need to overcome many bugs and issues, however, here are some existing working workflows that uses HandRefiner in ComfyUI:

Aoiryo commented 3 days ago

Thanks for your enquiry. Yeah looks like definitely something is wrong, what weight are you using, and what is the control strength? (it would be helpful if you can screenshot your entire workflow so I can try to debug)

I personally have never duplicated this project in ComfyUI, so if you try to duplicate it from scratch, it is very likely you need to overcome many bugs and issues, however, here are some existing working workflows that uses HandRefiner in ComfyUI:

Thanks for your timely response! I checked these workflows, and found that my results had weird edges cuz I didn't use 'set latent noise mask' node. Now I have the following workflow: 截屏2024-06-27 19 16 39 The hand shape is not as desired, is this because of my checkpoint or any wrong arguments?

wenquanlu commented 1 day ago

I am not 100% confidence of what's going on,

but from what I see from other workflows, in the "Load Checkpoint" section, they normally load a generic text-to-image weight (not an inpainting weight), so probably load this weight would work?

For example In this link https://openart.ai/workflows/datou/hand-refiner/XNcfT4iDYnQImq2UC3lP , they load SDXLAnimeBulldozer_v10.safetensors which is a text-to-image model In this link https://www.youtube.com/watch?v=K2uy2zcM68o , they load epicRealism.safetensors which is a text-to-image model In this link https://civitai.com/models/257359/awesomecomfyui-auto-hand-refiner-upscale-workflow-controlnet , they load realisian_v50.safetensors, which is a text-to-image model.

Please feel free to provide further feedback