Awesome work here! Overall it is working fantastically. It's pretty crazy the great results you can get in particular with close up portraits, even starting from low quality source images. When I generate images with a smaller face though the face is not as good and I think that's just an artifact of Stable Diffusion due to the compression in latent space. Typically I would fix this using something like FaceDetailer from https://github.com/ltdrdata/ComfyUI-Impact-Pack but I don't seem to be able to do that correctly.
What's interesting here is in the ApplyControlNetSEGS nodes they offer they don't actually accept positive/negative prompt and their full FaceDetailer expects clip directly. What I want to do is essentially mask the face and regenerate at higher resolution using the InstantID nodes, basically like ADetailer does. Any ideas of how that might be achieved?
manyart
commented
2 months ago
Awesome work here! Overall it is working fantastically. It's pretty crazy the great results you can get in particular with close up portraits, even starting from low quality source images. When I generate images with a smaller face though the face is not as good and I think that's just an artifact of Stable Diffusion due to the compression in latent space. Typically I would fix this using something like FaceDetailer from https://github.com/ltdrdata/ComfyUI-Impact-Pack but I don't seem to be able to do that correctly.
What's interesting here is in the ApplyControlNetSEGS nodes they offer they don't actually accept positive/negative prompt and their full FaceDetailer expects clip directly. What I want to do is essentially mask the face and regenerate at higher resolution using the InstantID nodes, basically like ADetailer does. Any ideas of how that might be achieved?