Could you please provide the model "1_mtrans_lossAllMaskNoMask" which was used to train Ctrlnet "cross" motion?

[Tomapua]

Could you please provide the model "1_mtrans_lossAllMaskNoMask" which was used to train Ctrlnet "cross" motion?

[ck10086]

Similar to the request above, I also hope to obtain the corresponding model version, or to find out whether there is relevant training code available in the repository？

[exitudio]

Hi,

TL;DR

• You can directly use the pretrained MoMask 2nd stage model.
• If you specifically need the 1_mtrans_lossAllMaskNoMask version, its weights are already included in the pretrained model, the Logits Regularization (ControlNet). But may need to manually separate it.

---

More context

The 1_mtrans_lossAllMaskNoMask is a modified version of the MoMask 2nd stage. Unlike the original MoMask, which computes the loss only on masked tokens, this version computes the loss on all tokens.

Because we need predictions for every token so they can be aggregated to obtain global positional information. So, I retrained the MoMask 2nd stage using a loss applied to all tokens.

However, after running a parameter search, I found that this modification provided little to no performance improvement. Since the pretrained model already incorporates this functionality during the Logits Regularization stage, I didn't to include this training in the repository to keep the codebase simpler.