How to restore a face in the correct way? - Githubissues

tencent-ailab / IP-Adapter

The image prompt adapter is designed to enable a pretrained text-to-image diffusion model to generate images with image prompt.

Apache License 2.0

4.46k stars 289 forks source link

How to restore a face in the correct way? #361

Open andapka opened 1 month ago

andapka commented 1 month ago

I'm using IP-Adapter + ControlNet with Normal maps:

IPAdapterPlus
base_model_path = "SG161222/Realistic_Vision_V6.0_B1_noVAE"
vae_model_path = "stabilityai/sd-vae-ft-mse"
image_encoder_path = "models/image_encoder/"
ip_ckpt = "models/ip-adapter-plus_sd15.bin" E.g. And I received a good picture, but in the photo, I don't look like myself at all.

I also tried ip_ckpt = "models/ip-adapter-plus-face_sd15.bin" the result is:

And tried face models, e.g.: ip_ckpt = "models/ip-adapter-faceid-plus_sd15.bin" my face extracted in a good way,

but the result is:

How can I achieve better results?

Appreciate your help

xiaohu2015 commented 1 month ago

do some high resolution fix

andapka commented 1 month ago

Should High resolution be applied to a face or the whole image?

Initially, I tried with image 1280x961 the result is:

In the above example, the image has a resolution of 3000x4000.