search

YuanxunLu / LiveSpeechPortraits

Live Speech Portraits: Real-Time Photorealistic Talking-Head Animation (SIGGRAPH Asia 2021)
MIT License
1.16k stars 200 forks source link

beta1 in train #21

Closed X-niper closed 2 years ago

X-niper commented 2 years ago

default bata1 in feature2face training is 0.5, while the article says you train your model with beta1 = 0.9. Is beta1 value critical in feature2face training?

Thanks!

YuanxunLu commented 2 years ago

Actually, I tried beta1 = 0.5 or 0.9 both in the early experiments, the results are similar. I think the influence of the beta1 value is not critical.

X-niper commented 2 years ago

Actually, I tried beta1 = 0.5 or 0.9 both in the early experiments, the results are similar. I think the influence of the beta1 value is not critical.

Thank you!!