Source
Bansal et al. (2022), LIT-553 — read as NOTE-299. Filed for
#163’s “[theory] cold diffusion”, and filed already retired, because the
evidence against it was in hand when it was filed (DP-003).
The account
Diffusion was explained through noise: Langevin dynamics, score matching, variational inference with a Gaussian prior. Cold Diffusion argued that these explanations were over-specific. Train any restorer on any continuous degradation, sample by restoring and re-degrading, and generation emerges. On this account the noise is one choice of degradation among many, and the theory built on it is “called into question”.
Why it is Rejected
The paper’s own generation numbers. On CelebA, noiseless cold generation from blur is FID 97.00 against 23.11 for the noise-based model. The authors trace the gap to a lack of diversity from perfectly correlated initial pixels, and fix it by adding Gaussian noise of σ = 0.002, which brings FID to 49.45. So a tiny amount of noise does half the work, and the rest of the gap remains.
The sweep that went looking. Warm Diffusion (LIT-555) holds the model and sampler fixed and moves only the blur-to-noise ratio. FID goes from 1.85 at 0.5 to 2.57 at 2 and 11.97 at 10. Generation quality tracks the amount of noise in the process, which is what this account says it should not do.
What survives
The procedure works conditionally. Deblurring, inpainting and super-resolution by iterated restoration improve FID over one-shot reconstruction. The sampler is exact for degradations linear in severity (LIT-553, §3.3). Neither needs this account. The replacement explanation is THEORY-080.