search

ali-vilab / UniAnimate

Code for Paper "UniAnimate: Taming Unified Video Diffusion Models for Consistent Human Image Animation".
https://unianimate.github.io/
1k stars 55 forks source link

Anime Faces do not look good #53

Open SWAPv1 opened 3 months ago

SWAPv1 commented 3 months ago

Hello.

Your examples show Chibi Anime characters with very clear face but when I run the inference it turns into semi realistic face with very low detail. Sometimes it just makes my Anime character which has a big head change to a human person.

Is there a setting I need to use such as resolution and steps and CFG to achieve a better result?

wangxiang1230 commented 3 months ago

Hi, there are some reasons for face inconsistency:

  1. The face of target pose is not aligned with the face of source image, probably because the anime face extracted by dwpose is not accurate.
  2. You could try using the first frame to generate the video, i.e. the ['image','local_image', "dwpose"] setting (https://github.com/ali-vilab/UniAnimate/blob/dfa7598b8368d22af9d8e89030220e3ae47d48ee/configs/UniAnimate_infer.yaml#L16)

And we are considering open sourcing an upgraded version of the model in the near future, which has a better appearance retention. Thanks for your attention.

SWAPv1 commented 3 months ago

Thank you for quick feedback. It is exciting to hear you will source a upgraded version!!