Feel free to raise issues!

[qinghew]

Please feel free to raise issues if you encounter any problems for your development. I will respond as soon as possible.

[qinghew]

If you tried before March 22 and found poor performance, please replace the base model with Stable Diffusion v2-1-base-512. Sorry for this bug.

[zengjie617789]

Thanks for the great contribution. I think more people prefer to use sdxl as base model though this work seems pretty well.

[qinghew]

Our work is more oriented towards academic research. Empirically, this method could be adapted to larger diffusion models.

[qinghew]

Since we use the ffhq-alignment tool to preprocess the input face to 512x512 and use "aligned_img = img_input.crop((64, 70, 440, 446))" to further crop:

StableIdentity/datasets_face/face_id.py

Line 153 in db9233d

aligned_img = img_input.crop((64, 70, 440, 446)) # Crop interesting region

.

If developers obtain bad results, please use the ffhq-alignment https://github.com/happy-jihye/FFHQ-Alignment to get aligned input face image.