search

OpenTalker / SadTalker

[CVPR 2023] SadTalker：Learning Realistic 3D Motion Coefficients for Stylized Audio-Driven Single Image Talking Face Animation
https://sadtalker.github.io/
Other
11.8k stars 2.2k forks source link

Face Enhancer Takes Most of the time, can it run in batches ? #478

Open mindbugg opened 1 year ago

mindbugg commented 1 year ago

current batch_size only works for face renderer, when running it on v100 GPU, most of the time is consumed in enhancer

TerragonDE commented 1 year ago

@mindbugg same here with a RTX4090... Face renderer takes a few Minutes, but Face Enhancer takes almost 2 hours and uses only 1 CPU Core !? ... Did you find a Solution?

akarce commented 1 year ago

I run the webui on Kaggle and it provides 2 x T4 GPU's but I couldn't find the optimal settings. batch size in generation is one of them. Additionally I need to utilize second GPU to get the best result bu IDK how...

jnc-nj commented 1 year ago

@mindbugg same here with a RTX4090... Face renderer takes a few Minutes, but Face Enhancer takes almost 2 hours and uses only 1 CPU Core !? ... Did you find a Solution?

Not really a 'found' solution per se, but you can rewrite the utils/face_enhancer.py file to include batching during face enhancing.

Since face enhancing only comes after face rendering, I just reused the batch_size argument for it.

snakeninny commented 1 year ago

Since face enhancing only comes after face rendering, I just reused the batch_size argument for it.

Do you mind to share your code?