Question about Cross-domain Att

[Ysz2022]

As shown in the appendix of the paper, it is said that attention contains: self-cross-domain att, row att, and cross-domain att. However, when I read the code, it seems that Only self-cross-domain att and row att are used, cross-domain att is discarded. Related details are as: Does Era3D actually only uses the self-cross-domain att and row att?

[pengHTYX]

Oh, this is a typo. Attention block in Era3D contains self-cross-domain att, row-wise MV att, and customer cross att. I will clarify this in the appendix.

[Ysz2022]

Could you please tell me the where is the "customer cross att" in the code? In which line of which file? Thank you very much !!!!

[pengHTYX]

https://github.com/pengHTYX/Era3D/blob/d6edf10b4b400a512a2c5be02960825e33b9b19d/mvdiffusion/models/transformer_mv2d_self_rowwise.py#L510C11-L510C12

[Ysz2022]

I have another 2 questions:
1、I completely agree with your row-wise attention. But for the self-cross-domain att, why you still compute the dense att instead of using the row-wise calculation?
2、For the customer cross att, it seems that this is a traditional cross-attention operation from the 'diffusers' repository, have you ever done any special designs?