Key techniques for maintaining facial consistency
Ways to prevent the generation of videos with distorted features:
- Model Selection:Confirm that it is loaded correctly
FantasyTalking_model.ckptcap (a poem)Wan2.1-I2V-14B-720Pbasic model - Focus Strengthening:Enable the built-in
facial attention cross-attentionmodule, add the--enable_fa_xattnparameters - Input specification:Use photos with even lighting on the front face to avoid interference such as reflections from glasses, and recommend using pure portraits after green screen keying.
- reprocessing program:For output video use
GFPGANPerform facial enhancement, example command:python enhance.py --input video.mp4 --type face
Key metrics: distortion is almost imperceptible to the human eye when the landmark offset is <5 pixels
This answer comes from the articleFantasyTalking: an open-source tool for generating realistic speaking portraitsThe
May not be reproduced without permission:AI productivity tools " How to avoid the problem of distorted facial features generated by FantasyTalking?
English 
简体中文 
日本語 
Deutsch 
Português do Brasil