DDIM makes diffusion sampling deterministic by setting the stochasticity parameter eta to zero, enabling far fewer steps

Denoising Diffusion Implicit Models (DDIM, Song et al. 2020) generalise the reverse process with a stochasticity parameter eta that scales the per-step noise variance. When eta=0, no noise is added at each reverse step and the process is fully deterministic; when eta=1, DDIM reduces to standard DDPM. The key benefit: the deterministic map lets you sample over only a subset of the timesteps. While DDPM needs the full T=1000 steps for good quality, DDIM with eta=0 achieves comparable quality in far fewer steps (e.g. tens) by traversing a sparser trajectory. This makes diffusion practical for interactive use. Determinism also gives a consistency property enabling semantically meaningful latent-space interpolation between samples.

Examples

DDIM with ~50 steps vs DDPM with 1000 steps: similar quality on CIFAR10 with roughly 20x fewer network evaluations. Semantic interpolation: encode two images to latents, interpolate, decode.

Assessment

What is the role of the eta parameter in DDIM? Contrast eta=0 with eta=1. Why can DDIM use far fewer sampling steps than DDPM without a large drop in quality?