Generation results look stylized?

OpenTalker / StyleHEAT

[ECCV 2022] StyleHEAT: A framework for high-resolution editable talking face generation

MIT License

620 stars 77 forks source link

Generation results look stylized? #39

Open shiliang26 opened 1 year ago

shiliang26 commented 1 year ago

Hi, I'm running the scripts of cross-id reenactment with provided demo images (pic 1). But the results I'm getting (pic 2) look suspiciously stylized. Is this simply expected results from styleGAN, or am I doing something wrong? Snipaste_2023-05-19_22-18-51

Thanks!

FeiiYin commented 1 year ago

Thanks for your attention. That is kind of an artifact of this framework. It mainly comes from the introduction of a post-training calibration network. The artifact will be alleviated a bit using optimizing inversion.

zhewei-mt commented 11 months ago

I am having the same problem. I tried same id and cross id, all results seem to have heavy artifact. However, the results video in https://feiiyin.github.io/StyleHEAT/ is more natural. Is there any configuration to be set to get natural results?