About Self-Conditioning

[v-shaoningli]

Hi all! Thanks for your excellent work RFdiffusion.

Here I have one question about the implementation of Self-Conditioning, i.e., Algorithm 3 in the supplementary.
Compare with the original paper Analog bits: Generating discrete data using diffusion models with self-conditioning, RFdiffusion additionally added ReverseStep to generate $x^{t}$ from $x^{t+1}$.

To be concrete, I attach the figure in the following:

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 $x^{pred}$.

Instead, in the original paper:

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

RFdiffusion has one additional step. Is it more benifitial to predict $x^{pred}$ with $x^{t}$ and $\tilde{x}_0$ compared with $x^{t+\Delta}$ and $\tilde{x}_0$ (because employing ReversteStep is a little bit more time-consuming)?

Looking forward to your responses. Thanks.

Best,
Shaoning

[KatarinaYuan]

Wondering the same thing here. Is this design better than the self-conditioning used in bit diffusion? How much is the performance gain?