TIFF: Text-Guided Generation of Full-Body Image with Preserved Reference Face for Customized Animation | Paper(ko), Paper(en), PPT
Python implementation
✨ This is an example of animated characters referencing Korean actors Park Eun-bin, Park Seo-joon, and Park So-yi. ✨
TIFFNet_Animation_Results.mov
- Performance comparison of generated images between existing methods and TIFF.
Method | L1↓ | Cos Dist.↓ | LPIPS↓ | SSIM↑ | PSNR↑ |
---|---|---|---|---|---|
LoRA | 1.169 | 0.845 | 0.750 | 0.426 | 11.106 |
InstantID | 1.064 | 0.694 | 0.762 | 0.368 | 10.666 |
TIFF(Ours) | 0.892 | 0.478 | 0.691 | 0.431 | 12.406 |
Ref Image(GT) | 0.838 | 0.419 | 0.386 | 0.546 | inf. |
- Performance comparison of generated videos between existing methods and TIFF.
Method | L1↓ | Cos Dist.↓ | LPIPS↓ | SSIM↑ | PSNR↑ |
---|---|---|---|---|---|
LoRA | 1.251 | 0.969 | 0.773 | 0.434 | 10.424 |
InstantID | 1.222 | 0.924 | 0.785 | 0.397 | 10.869 |
TIFF(Ours) | 0.967 | 0.568 | 0.723 | 0.438 | 12.293 |
Ref Video(GT) | 0.950 | 0.519 | 0.605 | 0.442 | 11.315 |
- The System Architecture for Generating Customized Animations with TIFF and Face Module.
- The Architecture of Champ Enhanced with the Face Module for Improved Facial Details.
For installation instructions, see the link below.
For Setup instructions, see the link below.
For Usage instructions, see the link below.
- This implementation was developed by Jaechan Jo
If you have any problem or error during running code, please email to us or submit an issue.