Research map / Image generation / Fast diffusion sampling
S$^2$-Guidance: Stochastic Self-Guidance for Training-Free Enhancement of Diffusion Models
Chubin Chen, Jiashu Zhu, Xiaokun Feng, Nisha Huang, Chen Zhu, Meiqi Wu, Fangyuan Mao, Jiahong Wu, Xiangxiang Chu, Xiu Li
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 S2-Guidance (stochastic block-dropping self-guidance) on top of CFG | SD3 · HPSv2.1 avg | score (%): 30.48 -> 31.09 | Improved |
| add S2-Guidance on top of CFG | SD3.5 · HPSv2.1 avg | score (%): 30.82 -> 31.56 | Improved |
| add S2-Guidance | SD3 · T2I-CompBench Color | score (%): 53.61 -> 59.63 | Improved |
| add S2-Guidance | SD3.5 · T2I-CompBench Color | score (%): 51.29 -> 57.57 | Mixed |
| add S2-Guidance | SiT-XL · ImageNet 256x256 FID | FID: 2.15 -> 2.03 | Improved |
| add S2-Guidance | Wan1.3B · VBench Total | score: 80.29 -> 80.93 | Improved |
| add S2-Guidance | Wan14B · VBench Total | score: 82.65 -> 82.84 | Improved |
| use APG guidance instead of CFG | Wan1.3B · VBench Total | score: 80.29 -> 70.83 | Hurt |
| compare S2-Guidance with CFG given equal compute (1.4x NFE) (ablation) | Flux de-distilled · HPSv2.1 avg | score (%): 30.13 -> 30.26 | Improved |
| drop more than ~3 of 24 blocks (ablation) | S2-Guidance SD3.5 · aesthetic score | Qalign: 4.652 -> 4.531 | Hurt |
| drop 1 block vs no dropping (ablation) | S2-Guidance SD3.5 · aesthetic score | Qalign: 4.618 -> 4.652 | Improved |
| use a single block-drop per step instead of naive multiple sub-networks (ablation) | S2-Guidance · aesthetic score | aesthetic score | No clear effect |
Later papers that compared with this method
1 later accepted paper reports S2 Guidance as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Improving Diffusion Generalization with Weak-to-Strong Segmented Guidance (CVPR 2026 · better on 4 of 4 shared results)
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
- Adaptive Guidance: Training-free Acceleration of Conditional Diffusion Models (AAAI 2025 · similarity 0.93)
- No Training, No Problem: Rethinking Classifier-Free Guidance for Diffusion Models (ICLR 2025 · similarity 0.92)
- Eliminating Oversaturation and Artifacts of High Guidance Scales in Diffusion Models (ICLR 2025 · similarity 0.92)
- Rectified Diffusion Guidance for Conditional Generation (CVPR 2025 · similarity 0.92)
- TFG: Unified Training-Free Guidance for Diffusion Models (NeurIPS 2024 · similarity 0.91)
- Characteristic Guidance: Non-linear Correction for Diffusion Model at Large Guidance Scale (ICML 2024 · similarity 0.91)
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.
