search

Dai-Wenxun / MotionLCM

[ECCV 2024] MotionLCM: This repo is the official implementation of "MotionLCM: Real-time Controllable Motion Generation via Latent Consistency Model"
Other
176 stars 5 forks source link

About pretrained MLD #11

Closed DunyangNi closed 1 week ago

DunyangNi commented 2 weeks ago

Is it possible to train MotionLCM on a checkpoint pretrained with the original MLD repo? If not, what are the configurations you use to pretrain the MLD?

Dai-Wenxun commented 2 weeks ago

If I remember correctly, we train MLD on the original MLD repo and then directly load the pre-trained checkpoint to train MotionLCM. But there are some useless attributes in the pre-trained checkpoint as shown below. You only need to remain the 'state_dict'.

image

The LAST step is to delete one useless parameter saved in 'denoiser', 'denoiser.mem_pos.pe'. This param is not been used in denoiser but the author left it carelessly.

Finally, save the cleaned checkpoint which is a dict with ONE key state_dict.

If you encounter any issues in this process, feel free to inform me. hahaha

Dai-Wenxun commented 1 week ago

Feel free to open this issue