Research line · emerging

Natively few-step objectives

Train for one or few steps from the start, instead of training a many-step model and distilling it afterwards.

What defines membership

Iterative sampling is an artifact of the training objective, and a better objective yields one-step generation directly.

Read top to bottom: what came before, the idea itself, the evidence for it, what improved, and where it breaks.

Conditions the network on step size so one model works at any budget down to a single step, with no distillation stage.

IMM · 2025-03core

Matches distributions between interpolant marginals in one training stage, with convergence guarantees consistency models lack; 1.99 FID in 8 steps.

MeanFlow · 2025-05landmark

Targets average rather than instantaneous velocity, giving a self-contained identity for from-scratch one-step training; 3.43 FID at 1 NFE.

SANA-Sprint · 2025-03strong-followup

The distillation-based contrast case: 1-4 steps via continuous-time consistency distillation plus an adversarial term, but requires a teacher.

Drifting Models · 2026-02landmark

Learns the generator mapping directly and evolves its pushforward until it matches the data; natively one-step, 1.54 FID on ImageNet 256.

Pixel MeanFlow · 2026-01strong-followup

Combines one-step training with no tokenizer at all.

Applies drifting-style distillation inside RAE latents, where anisotropy makes the naive version unstable.

Collapses the inference-cost gap that made diffusion expensive to serve
No teacher model, so no inherited teacher errors and one training run instead of two
Advancing fast, with several independent formulations converging on similar quality

Still generally behind the best many-step models at the very top of the quality range
Mostly demonstrated on class-conditional ImageNet rather than frontier text-to-image
Interacts badly with high-dimensional anisotropic latents, as the RAE case shows

As published, with the guidance method, budget, and model size that produced them. Seethe comparison page for why these cannot be ranked naively.

Paper	Benchmark	Value	Guidance	Budget	Params
Distilling Drifting Transformers	ImageNet 256x256 gFID (NFE=1)	1.48	extrapolation-based	16 epochs distillation	839M (inherited from RAE)
Drifting Models	ImageNet 256x256 gFID (NFE=1)	1.61	not stated	not stated	not stated
Drifting Models	ImageNet 256x256 gFID (NFE=1)	1.54	not stated	not stated	not stated
IMM	ImageNet 256x256 gFID (NFE=8)	1.99	not stated	not stated	not stated
MeanFlow	ImageNet 256x256 gFID (NFE=1)	3.43	not stated	not stated	not stated