Research map / Image generation / Fast diffusion sampling
SANA-Sprint: One-Step Diffusion with Continuous-Time Consistency Distillation
Junsong Chen, Shuchen Xue, Yuyang Zhao, Jincheng Yu, Sayak Paul, Junyu Chen, Han Cai, Song Han, Enze Xie
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add LADD adversarial loss to sCM (ablation) | SANA-Sprint · MJHQ-30K 2-step | FID: 8.93 -> 8.11 | Improved |
| add LADD adversarial loss to sCM (ablation) | SANA-Sprint · MJHQ-30K 2-step | CLIP score: 27.51 -> 28.02 | Improved |
| use LADD alone instead of sCM plus LADD (ablation) | SANA-Sprint · MJHQ-30K 2-step | FID: 8.11 -> 12.20 | Hurt |
| add CFG embedding to student (ablation) | SANA-Sprint · MJHQ-30K 2-step | CLIP score: 27.15 -> 28.09 | Improved |
| add CFG embedding to student (ablation) | SANA-Sprint · MJHQ-30K 2-step | FID: 9.23 -> 8.72 | Improved |
| add max-time weighting at t=pi/2 (50%) (ablation) | SANA-Sprint · MJHQ-30K 2-step | FID: 9.44 -> 8.32 | Improved |
| raise max-time probability from 50% to 70% (ablation) | SANA-Sprint · MJHQ-30K 2-step | FID: 8.32 -> 8.11 | No clear effect |
| sCM:LADD weight 1.0:0.5 vs 1.0:1.0 (ablation) | SANA-Sprint · MJHQ-30K 2-step | FID: 8.81 -> 8.43 | Improved |
| sCM:LADD weight 1.0:0.1 vs 1.0:0.5 (ablation) | SANA-Sprint · MJHQ-30K 2-step | FID: 8.43 -> 8.90 | Hurt |
| training-free flow to TrigFlow transformation | SANA · MJHQ-30K 50-step Euler | FID: 5.81 -> 5.73 | No clear effect |
| use schedule transfer when fine-tuning flow model to TrigFlow (ablation) | SANA · fine-tuning | training stability | Improved |
| set c_noise(t)=t instead of 1000t (dense time embedding) (ablation) | SANA teacher · gradient norm | gradient norm: exceeding 10^3 -> more stable levels | Improved |
Later papers that compared with this method
2 later accepted papers report SANA-Sprint 0.6B as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- SD3.5-Flash: Distribution-Guided Distillation of Generative Flows (ECCV 2026 · better on 0 of 4 shared results)
- InverFill: One-Step Inversion for Enhanced Few-Step Diffusion Inpainting (CVPR 2026 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Score identity Distillation: Exponentially Fast Distillation of Pretrained Diffusion Models for One-Step Generation (ICML 2024 · similarity 0.85)
- Random Conditioning with Distillation for Data-Efficient Diffusion Model Compression (CVPR 2025 · similarity 0.84)
- Diffusion Model is Effectively Its Own Teacher (CVPR 2025 · similarity 0.84)
- Adaptive Stochastic Coefficients for Accelerating Diffusion Sampling (NeurIPS 2025 · similarity 0.84)
- SCott: Accelerating Diffusion Models with Stochastic Consistency Distillation (AAAI 2025 · similarity 0.84)
- Continuous-Time Distribution Matching for Few-Step Diffusion Distillation (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
