驱动视频如果比较糊，会导致生成视频也比较糊吗？尤其是牙齿细节上

[sixone-Jiang]

[huangxin168]

实测过，应该是有影响的

[cleardusk]

Thanks for your report. @sixone-Jiang @huangxin168 Could you please provide an example, e.g., provide side-by-side comparisons ?

[sixone-Jiang]

很抱歉我不能提供一个视频，但我可以给你提供我实验时的核心设置， 'flag_normalize_lip': False, 'flag_relative_motion': False, 'flag_do_crop': True, 'flag_pasteback': True, 'flag_stitching': True, 'animation_region': "lip", 'driving_option': "expression-friendly", 对于driving vedio，可以先对人脸下采样两倍，采取kernel=5的高斯模糊，再上采样到原分辨率用以获取低分辨率

[sixone-Jiang]

更进一步的，我发现对于厚嘴唇的样例（source video），会较难让嘴唇有足够的张合大小

[cleardusk]

很抱歉我不能提供一个视频，但我可以给你提供我实验时的核心设置， 'flag_normalize_lip': False, 'flag_relative_motion': False, 'flag_do_crop': True, 'flag_pasteback': True, 'flag_stitching': True, 'animation_region': "lip", 'driving_option': "expression-friendly", 对于driving vedio，可以先对人脸下采样两倍，采取kernel=5的高斯模糊，再上采样到原分辨率用以获取低分辨率

It seems that the motion extractor is sensitive to the disturb of driving images/video. It's a good point. A future solution is to perform augmentation (e.g., blur) on driving images when training the motion extractor.