Tencent / MimicMotion

High-Quality Human Motion Video Generation with Confidence-aware Pose Guidance
https://tencent.github.io/MimicMotion/
Other
1.93k stars 165 forks source link

Optimize VRAM Usage. Works on 16G 4060ti now! #32

Closed zyayoung closed 4 months ago

zyayoung commented 4 months ago

This pull request implements:

The pipeline requires only 16G VRAM (tested on 4060ti) and finishes in 20 minutes on a 4090 GPU for the 35s demo using our 72-frame model. This information answers https://github.com/Tencent/MimicMotion/issues/17 and is added to the Readme. This fix should also resolve the runtime issue in https://github.com/Tencent/MimicMotion/issues/27 caused by VRAM spilling into slow CPU memory.

Minor change: Avoid downloading the entire SVD-1.1 repository; instead, download only the necessary files during execution.