[🎉ICLR2026] S²-Guidance: Stochastic Self-Guidance for Training-Free Enhancement of Diffusion Models

Chubin Chen1, Jiashu Zhu2, Xiaokun Feng2,3, Nisha Huang1 Meiqi Wu2,3, Fangyuan Mao2, Jiahong Wu2,‡, Xiangxiang Chu2, Xiu Li1,† 1Tsinghua University 2AMAP, Alibaba Group 3CASIA †Corresponding author. ‡Project lead.

</div>

🔥 Updates

[2025/08] Our paper is available on arXiv and the project page is live!
[2026/01]: 🎉🎉🎉 Our paper is accepted by ICLR 2026!
[Coming Soon] Code will be released in few days!

📣 Overview

Method Overview Diagram

We propose S²-Guidance, a novel method that leverages stochastic block-dropping during the forward process to construct sub-networks, effectively guiding the model away from potential low-quality predictions and toward high-quality outputs. Extensive qualitative and quantitative experiments on text-to-image and text-to-video generation tasks demonstrate that S²-Guidance delivers superior performance, consistently surpassing CFG and other advanced guidance strategies.

🎉 Results

Here are some examples comparing the results from standard Classifier-free Guidance (CFG) with our S²-Guidance. Our method consistently produces higher-fidelity and more coherent results.

Text-to-Image Generation

Prompt: "The bold dramatic strokes of the painter's brush created a stunning abstract masterpiece a work of emotional depth and intensity."
| CFG | Ours (S²-Guidance) |
|:--------------:|:------------------:|
| <img src="assets/baseline1.jpg" alt="CFG Image 1"> | <img src="assets/ours1.jpg" alt="Ours Image 1"> |

Prompt: "A floating castle above the clouds, with 'S2 Guidance Is All You Need' swirling in the mist."
| CFG | Ours (S²-Guidance) |
|:--------------:|:------------------:|
| <img src="assets/baseline2.jpg" alt="CFG Image 2"> | <img src="assets/ours2.jpg" alt="Ours Image 2"> |

Prompt: "A woman is holding a bouquet of balloons and celebrating a birthday."
| CFG | Ours (S²-Guidance) |
|:--------------:|:------------------:|
| <img src="assets/baseline3.jpg" alt="CFG Image 3"> | <img src="assets/ours3.jpg" alt="Ours Image 3"> |

Prompt: "A red book and an ivory sheep."
| CFG | Ours (S²-Guidance) |
|:--------------:|:------------------:|
| <img src="assets/baseline4.jpg" alt="CFG Image 4"> | <img src="assets/ours4.jpg" alt="Ours Image 4"> |

Prompt: "A cat sitting besides a rocket on a planet with a lot of cactuses."
| CFG | Ours (S²-Guidance) |
|:--------------:|:------------------:|
| <img src="assets/baseline5.jpg" alt="CFG Image 5"> | <img src="assets/ours5.jpg" alt="Ours Image 5"> |

Prompt: "A woman sitting under an umbrella in the middle of a restaurant."
| CFG | Ours (S²-Guidance) |
|:--------------:|:------------------:|
| <img src="assets/baseline6.jpg" alt="CFG Image 6"> | <img src="assets/ours6.jpg" alt="Ours Image 6"> |

Text-to-Video Generation

Prompt: "A breathtaking close-up of a woman frozen in time as golden threads of light weave around her face, creating dynamic flowing patterns of energy amidst glowing particles."
<table border="0" style="width: 100%; text-align: center; margin-top: 20px;">
<tr>
<td style="text-align: center;">
<b>CFG</b>
<br>
<video src="https://github.com/user-attachments/assets/ca181aed-1889-4b90-a885-ef429dc861e0" width="100%" controls autoplay loop muted playsinline alt="CFG Video 4"></video>
</td>
<td style="text-align: center;">
<b>Ours (S²-Guidance)</b>
<br>
<video src="https://github.com/user-attachments/assets/b14c92be-860a-4596-91f3-ed8ce4e0e9cd" width="100%" controls autoplay loop muted playsinline alt="Ours Video 4"></video>
</td>
</tr>
</table>

Prompt: "An astronaut flying in space."
<table border="0" style="width: 100%; text-align: center; margin-top: 20px;">
<tr>
<td style="text-align: center;">
<b>CFG</b>
<br>
<video src="https://github.com/user-attachments/assets/255444a3-3130-4747-913c-fc9ad3e11ae8" width="100%" controls autoplay loop muted playsinline alt="CFG Video 3"></video>
</td>
<td style="text-align: center;">
<b>Ours (S²-Guidance)</b>
<br>
<video src="https://github.com/user-attachments/assets/067fba6c-c244-4f6a-a55f-5ba785ad6264" width="100%" controls autoplay loop muted playsinline alt="Ours Video 3"></video>
</td>
</tr>
</table>

Prompt: "A car accelerating to gain speed."
<table border="0" style="width: 100%; text-align: center; margin-top: 20px;">
<tr>
<td style="text-align: center;">
<b>CFG</b>
<br>
<video src="https://github.com/user-attachments/assets/8b1903e6-d45c-4411-81f7-4d93a008c427" width="100%" controls autoplay loop muted playsinline alt="CFG Video 2"></video>
</td>
<td style="text-align: center;">
<b>Ours (S²-Guidance)</b>
<br>
<video src="https://github.com/user-attachments/assets/0886e3a1-e60e-49b3-9b7a-8b300820892d" width="100%" controls autoplay loop muted playsinline alt="Ours Video 2"></video>
</td>
</tr>
</table>

Prompt: "A close-up of a beautiful woman's face with colored powder exploding around her, creating an abstract splash of vibrant hues."
<table border="0" style="width: 100%; text-align: center; margin-top: 20px;">
<tr>
<td style="text-align: center;">
<b>CFG</b>
<br>
<video src="https://github.com/user-attachments/assets/d8581bb7-a224-4dcf-a71d-3e3afd6aa8e3" width="100%" controls autoplay loop muted playsinline alt="CFG Video 1"></video>
</td>
<td style="text-align: center;">
<b>Ours (S²-Guidance)</b>
<br>
<video src="https://github.com/user-attachments/assets/cf891170-6c03-486c-86db-d171707d5fe8" width="100%" controls autoplay loop muted playsinline alt="Ours Video 1"></video>
</td>
</tr>
</table>

User Study

Quantitative Results

HPSV2.1 & T2I-ComBench

VBench

🙏 Acknowledgements

This work is built upon many amazing open-source projects. We would like to thank the developers of Diffusers, PyTorch, and other related libraries for their contributions to the community.

📜 Citation

If you find our work useful for your research, please feel free to leave a star⭐️⭐️⭐️ and cite our paper:

@inproceedings{chenstochastic,
  title={Stochastic Self-Guidance for Training-Free Enhancement of Diffusion Models},
  author={Chen, Chubin and Zhu, Jiashu and Feng, Xiaokun and Huang, Nisha and Zhu, Chen and Wu, Meiqi and Mao, Fangyuan and Wu, Jiahong and Chu, Xiangxiang and Li, Xiu},
  booktitle={The Fourteenth International Conference on Learning Representations}
}

S2-Guidance

README