Research map / Image generation / Text-to-image diffusion
SnapFusion: Text-to-Image Diffusion Model on Mobile Devices within Two Seconds
Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, Jian Ren
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 |
|---|---|---|---|
| use CFG-aware step distillation loss alongside vanilla distillation (ablation) | 8-step efficient UNet · MS-COCO 2014 val | CLIP score: +0.002 to +0.007 CLIP over SD-v1.5 at similar FID | Improved |
| vanilla step distillation only (no CFG-aware loss) (ablation) | 8-step student · MS-COCO 2014 val | CLIP score: capped at 0.30 | Hurt |
| mix vanilla and CFG distillation losses (ablation) | step distillation · MS-COCO 2014 val 6K | FID and CLIP | Improved |
| direct distillation from 32-step teacher instead of progressive distillation (ablation) | step distillation of SD-v1.5 · MS-COCO 2014 val 6K | FID and CLIP | Improved |
| add original denoising loss during CFG distillation (ablation) | CFG distillation · MS-COCO 2014 val 6K | FID | Improved |
| use more teacher steps (32/64 instead of 16) for 8-step student (ablation) | vanilla step distillation · MS-COCO 2014 val 6K | CLIP score | Hurt |
| increase CFG probability and CFG range (ablation) | CFG distillation · MS-COCO 2014 val 6K | FID vs CLIP | Mixed |
| add robust training (stochastic block skipping) before evolving architecture (ablation) | SD-v1.5 UNet · qualitative block removal | image quality after removing blocks | Improved |
| distill to 8 steps with CFG-aware loss vs w-conditioned distillation at 16 steps (ablation) | SD-v1.5 distillation · MS-COCO 2014 val 6K | FID and CLIP | Improved |
| reduce UNet steps from 16 to 8 without distillation (ablation) | efficient UNet · MS-COCO 2014 val 6K | FID and CLIP | Hurt |
| apply CFG-aware distillation to SD-v2 for 8 steps | SD-v2 · MS-COCO 2014 val | FID and CLIP | No clear effect |
Later papers that compared with this method
2 later accepted papers report SnapFusion as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- Clockwork Diffusion: Efficient Generation With Model-Step Distillation (CVPR 2024 · better on 2 of 2 shared results)
- MobileDiffusion: Instant Text-to-Image Generation on Mobile Devices (ECCV 2024 · 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
- SnapGen: Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training (CVPR 2025 · similarity 0.92)
- Würstchen: An Efficient Architecture for Large-Scale Text-to-Image Diffusion Models (ICLR 2024 · similarity 0.90)
- Ambient Diffusion Omni: Training Good Models with Bad Data (NeurIPS 2025 · similarity 0.89)
- InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation (ICLR 2024 · similarity 0.89)
- On Discrete Prompt Optimization for Diffusion Models (ICML 2024 · similarity 0.89)
- Shielded Diffusion: Generating Novel and Diverse Images using Sparse Repellency (ICML 2025 · similarity 0.88)
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.
