search

TMElyralab / MuseTalk

MuseTalk: Real-Time High Quality Lip Synchorization with Latent Space Inpainting
Other
2.51k stars 306 forks source link

开源的唇语驱动模型，是从随机初始化开始训练的，还是先对Unet网络结构进行预训练后再训练唇语驱动模型呢？ #95

Closed gobigrassland closed 4 months ago

gobigrassland commented 4 months ago

我看到用到的Unet模型参数与SD1.4模型配置参数，就是其中cross_attention_dim和in_channels的区别。 （1）唇语模型UNet： cross_attention_dim=384, in_channels=8 （2）SD1.4 UNet: cross_attention_dim=768, in_channels=4

czk32611 commented 4 months ago

是从随机初始化开始训练的