Few-Step Distillation of Diffusion Models
Compressing slow diffusion generators into models that produce an image or video in one to eight steps.
Tier Grades
Tier A: Frontier AI labs
feasible
Big labs release distilled variants with each model.
Tier B: Indonesian AI labs
feasible
Distilling an open model needs far less compute than pretraining it, so it fits an Indonesian lab.
Tier C: Institution units such as ITB
feasible
A 3-author team produced SDXL-Lightning; a university unit can do this.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
feasible
Few-step models cut serving cost in generation pilots.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeTraining compute is not in the sources read.
- DataNot in the sources read; SDXL-Lightning distils SDXL for one-step and few-step 1024px generation.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 3 authors on the SDXL-Lightning paper (arXiv 2402.13929).
- Time horizonMonths.
- Shelf-life signalsOne-step generation matches full-step quality, Distilled turbo variants ship with every major release
- Pipeline roleTier A releases distilled models; Tier C studies distillation; Tier D uses fast variants to cut cost in pilots.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
Observed actors
Linked models
FLUX.1 [schnell]image generationexemplar
SDXL-Lightningimage generationexemplar
SD-Turboimage generationexemplar
Latent Consistency Models (LCM)image generationexemplar
Krea 2 Turboimage generationproduced byFLUX.2 [klein] 9Bimage generationproduced bySeedVR2video generationproduced by
SANA-Sprint 1.6Bimage generationproduced byAnimateDiff-Lightningvideo generationproduced by
AnimateLCMvideo generationproduced byLCM Dreamshaper v7image generationproduced byLinks
R&D TopicFew-Step Distillation of Diffusion Models
ModelsFLUX.1 [schnell]Black Forest Labs, image generationSD-TurboStability AI, image generationSDXL-LightningByteDance, image generationLatent Consistency Models (LCM)Tsinghua University, image generation+7 more
FLUX.1 [schnell]Black Forest Labs, image generationSD-TurboStability AI, image generationSDXL-LightningByteDance, image generationLatent Consistency Models (LCM)Tsinghua University, image generation+7 moreUse Casesnone recorded
Who Runs Itnone recorded
Models 11
AnimateLCMThe Chinese University of Hong Kong, SenseTime, video generationproduced byKrea 2 TurboKrea.ai, Inc., image generationproduced byAnimateDiff-LightningByteDance, video generationproduced byFLUX.1 [schnell]Black Forest Labs, image generationexemplarFLUX.2 [klein] 9BBlack Forest Labs, image generationproduced bySeedVR2ByteDance, video generationproduced byLatent Consistency Models (LCM)Tsinghua University, image generationexemplarLCM Dreamshaper v7Tsinghua University, image generationproduced bySANA-Sprint 1.6BNVIDIA, image generationproduced bySD-TurboStability AI, image generationexemplarSDXL-LightningByteDance, image generationexemplarPapers
- Progressive Distillation for Fast Sampling of Diffusion ModelsTim Salimans et al., ICLR 2022 (per arXiv comment), 2022 founding
- Consistency ModelsYang Song et al., ICML 2023 (per arXiv comment), 2023 founding
- Improved Distribution Matching Distillation for Fast Image SynthesisTianwei Yin et al., arXiv preprint, 2024 state of the art
- Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step InferenceSimian Luo et al., arXiv preprint, 2023 state of the art
- Adversarial Diffusion DistillationAxel Sauer et al., arXiv preprint, 2023 state of the art
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- arxiv.org/abs/2402.13929 https://arxiv.org/abs/2402.13929
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Few-Step Distillation of Diffusion Models", https://aurion.semai.id/topics/few-step-diffusion-distillation/, records of October 2026.