Details of the 72-frame Model

Tencent / MimicMotion

High-Quality Human Motion Video Generation with Confidence-aware Pose Guidance

https://tencent.github.io/MimicMotion/

Other

1.95k stars 166 forks source link

Details of the 72-frame Model #52

Open zhysora opened 4 months ago

zhysora commented 4 months ago

Thanks for your great works! Could you please share some details of the 72-frame model. How to train such a big model. It seems to require pretty large GPU Memory. Do you still train the full params in SVD, or just some temporal related layers, or use some other training tricks. Thank you!

andreclaudino commented 4 months ago

I would love to know more about s possible fine tunning process. Are you able to share a fine tune tutorial?