Facial anomalies prevention and repair program
Facial distortion is usually caused by input quality issues or model overfitting, and the following countermeasures are recommended:
- Enter normative prevention::
- The video needs to include the front face and be unobstructed (OpenCV auto-detection is recommended)
- Audio length does not deviate from video length by more than ±10%
- Solid colors are recommended for the background (easy matting)
- Parameter tuning program::
- increase
--pose_style 0Limit head movement - align
--expression_scale 0.5Reduced intensity of expression - Try different pre-trained models (especially download the Baidu.com disk of
exp_model(Expansion pack)
- increase
- Repair of reprocessing::
- Using the Deformation Stabilizer in Adobe After Effects
- Localized corrections via Face2Face tool
- Keyframe manual adjustment (set correction points every 5 frames)
For special cases, try theinference.pyAdjustment near line 127face3d_helperof the mesh smoothing parameter, with recommended values ranging from 0.3 to 0.7. CUDA/cuDNN version compatibility should be checked for severe distortions.
This answer comes from the articleSVLS: SadTalker Enhanced to Generate Digital People Using Portrait VideoThe
May not be reproduced without permission:AI productivity tools " How to avoid abnormalities such as facial distortion in SVLS generated videos?
