Can consistency models trade speed for quality with a few steps?

Diffusion models are easy to train but slow to sample (many function evaluations); consistency models sample in a single step but are hard to train and sacrifice quality. Multistep Consistency Models unify Consistency Models and TRACT into a single dial: a 1-step model is a conventional consistency model, while an ∞-step model is a diffusion model — so the method interpolates between the two. The practical finding is that a small budget increase (1 → 2–8 steps) makes models much easier to train and yields higher-quality samples while retaining most of the single-step speed advantage — closing the quality gap to standard diffusion in as few as 8 steps, and scaling to text-to-image.

The keeper is the framing of sampling-steps as a continuous quality-speed trade-off rather than a binary choice between "fast but worse" (consistency) and "slow but best" (diffusion). The hard single-step regime was the wrong target; a handful of steps recovers most quality at most of the speed.

This sits in the vault's diffusion thread as a sampling-efficiency contribution. It pairs with Can generating entire videos at once beat keyframe interpolation? (Lumiere) as another rethinking of the diffusion generation budget, and the speed-quality dial mirrors the test-time-compute trade-offs seen on the language side.

Easy Consistency Tuning makes the diffusion→consistency conversion cheap (ECT, https://arxiv.org/abs/2406.14548). The same "diffusion is a special case of consistency" view powers a training-efficiency result: rather than training a consistency model from scratch (a week on 8 GPUs as of 2024), ECT fine-tunes a pretrained diffusion model and progressively approximates the full consistency condition over training. It reaches a 2-step FID of 2.73 on CIFAR-10 in ~1 hour on a single A100 — matching Consistency Distillation that took hundreds of GPU-hours — and the resulting consistency models obey classic power-law scaling, suggesting they improve with scale. So Multistep Consistency dials inference steps along the diffusion↔consistency continuum; ECT exploits the same continuum to make training the consistency model cheap by starting from diffusion. (ECT's one limitation: it needs the dataset, unlike data-free distillation.)