jaechanjo / TIFF

Text-Guided Generation of Full-Body Image with Preserved Reference Face for Customized Animation

TIFF

TIFF: Text-Guided Generation of Full-Body Image with Preserved Reference Face for Customized Animation | Paper(ko), Paper(en), PPT

Python implementation

Jaechan Jo

✨ This is an example of animated characters referencing Korean actors Park Eun-bin, Park Seo-joon, and Park So-yi. ✨

TIFFNet_Animation_Results.mov

Performace by method

1. 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.

2. 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

Method Overview

1. The System Architecture for Generating Customized Animations with TIFF and Face Module.

2. The architecture of proposed TIFF and Face Module.

3. The Architecture of Champ Enhanced with the Face Module for Improved Facial Details.

Evaluation

For installation instructions, see the link below.

HowToEval

Setup

For Setup instructions, see the link below.

HowToSetup

Usage

For Usage instructions, see the link below.

HowToUse

Others

• 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.