trends.barnas.net

method · first seen 9 Oct 2026

few-step diffusion sampling

Also: few-step sampling, accelerated sampling, few-step diffusion models, few-step generative models

Techniques that reduce the number of model evaluations needed to generate samples from diffusion or flow generative models.

Few-step diffusion sampling covers methods that cut the number of network evaluations needed to draw samples from diffusion and flow models, through distillation, better solvers, or interval-average velocity models such as MeanFlow. The sources show steady incremental work rather than a new idea. Papers correct endpoint errors after training (TraceSampler), separate prediction horizons from state updates in MeanFlow, and extend few-step sampling to masked diffusion language models such as LLaDA-8B. Several groups, including one with Jong Chul Ye, adapt reward alignment and reinforcement learning to few-step generators. Code releases cover velocity adversarial distillation and RL fine-tuning on SD-3.5.

Sampling cost has been a central diffusion research problem since DDIM, consistency models and progressive distillation, so this activity reflects an established area rather than a new trend. The newer threads worth tracking are RL and reward alignment for few-step models and few-step sampling for masked diffusion language models.

Analyst note (established), 17 Nov 2026.

Mentions per week

0419 May11 Aug10 NovPapersCodeDiscussion

PapersCodeDiscussion

Rank 49 on 14 Nov 2026 · 6 mentions in 28 days · 6 independent groups · papers → code · first seen 5 weeks ago · acceleration +2.02

Earliest sources

Latest sources

Mentioned together with

rectified flowlearned residual correction for samplingflow matchingmasked diffusion modelreward-tilted samplingtrajectory distillationdiffusion language modelinitial noise optimizationvelocity adversarial distillationreinforcement learning for diffusion models