Cold Diffusion: Inverting Arbitrary Image Transforms Without Noise

Interesting. After reading this, I suspect there must be some kind of universal principle at work when we iteratively degrade a sample x with severity t and then generate/recover x from the degraded state, again and again, decreasing t in each iteration -- as shown in Algorithm 2 of this paper.

Specifically, given that Algorithm 2 works for both non-deterministic and deterministic degradations (i.e., both with and without adding noise), I'm not sure we can properly call the process "diffusion" anymore. When the degradation is deterministic, there's no hot state, properly speaking, from which to anneal into a cold state.

If I understand this correctly, diffusion appears to be a special case of a more general iterative process. Or am I missing something here?