Is there a non-stochastic sampling method?

[krNeko9t]

I want to directly use CM to get the whole trajectory of xT -> x0, like an ODE sovler, is it possible?

[ljw919]

care about this too

[PeterL1n]

It is possible.

Because the model is trained to support any x_t -> x_0, in theory you can always use the same noise to renoise x_0 on every step, which will get you a deterministic output. However, in practice it doesn't work very well and will have a much larger error accumulation.

I have found that the problem is mainly due to the network inable to predict very clean x_0 on every step. Because x_t = diffuse(x_0, eps). The network is essentially predicting diffuse(x_0, eps) -> x_0 on every step.The network predicted x_0 is prone to have some residual of eps, and if we use the same eps to renoise in the next step, the eps compounds and causes the whole statistics to shift. Using a different eps mitigates this compounding issue.

So for practical purposes, if you want deterministic results, you can cheat by just setting a seed when generating all random noises for intermediate step. Or re-use the same noise but do some transformation, like flipping, rotating90deg, etc to avoid exact error compounding.