Source
Song et al. (2020), LIT-038 — DDIM.
The claim
The sampler family LIT-038 derives is parameterised by how much noise it injects at each step. At zero, the generative process is a deterministic map from the initial noise to the sample, and two capabilities appear that a stochastic sampler structurally cannot have:
- The initial noise determines high-level image content regardless of how many steps you take. It therefore behaves as a latent code, and interpolating in it produces semantically meaningful interpolation.
- Samples can be encoded back to their latent and reconstructed with very low error.
DDPM cannot do either, and the reason is not a tuning gap: its sampling adds fresh noise at every step, so the trajectory does not carry the starting point forward.
When this applies
Reach for it when the latent is part of the product — editing, inversion, interpolation, or anything that needs the same input to give the same output.
It is not a blanket recommendation to sample deterministically. LIT-075 finds that the useful amount of stochasticity depends on how good the model is: noise injected during sampling corrects accumulated error, and a weaker model benefits from it. Determinism buys invertibility and costs that correction.
Conditions
- 2020, images, DDPM-scale models.
- Deterministic trajectories accumulate error at the low signal-to-noise end. LIT-038 does not examine this; LIT-067 is where it has to be fixed.
- Determinism here is a property of the sampler, not of the parametrization. It composes with SOTA-188 and SOTA-195 rather than competing with them.