How to do self-conditioning?

[KatarinaYuan]

Hi, I noticed that in the paper, you mentioned that self-conditioning is quite important for RFdiffusion. Wondering if you could provide an example of implementation for us to follow? Thank you!

[v-shaoningli]

What RFdiffusion is done:

1. Generate $x^{t+\Delta}$;
2. Employ ReversteStep to generate $x^{t}$;
3. Use RFdiffusion to predict $\tilde{x}_0$ and stop gradient;
4. Use $x^{t}$ and $\tilde{x}_0$ as input of RFdiffusion to predict $x^{pred}$.

[KatarinaYuan]

@v-shaoningli created a discussion (https://github.com/RosettaCommons/RFdiffusion/discussions/98). I think we wonder similar questions: (1) whether the simplified implementation adopted by bit diffusion would hurt the performance for RFDiffusion (2) whether this is the reason RFDiffusion chooses this more complicated way of implementation and how much performance loss it would cause

Thank you in advance!