CCS: Controllable and Constrained Sampling with Diffusion Models via Initial Noise Perturbation
Bowen Song, Zecheng Zhang, Zhaoxu Luo, Jason Hu, Wei Yuan, Jing Jia, Zhengxu Tang, Guanyang Wang, Liyue Shen · 2025
Read CCS: Controllable and Constrained Sampling with Diffusion Models via Initial Noise Perturbation by Bowen Song, Zecheng Zhang, Zhaoxu Luo, Jason Hu, Wei Yuan, Jing Jia, Zhengxu Tang, Guanyang Wang, Liyue Shen (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-04670.txt.
Description
Bowen Song and eight co-authors find a simple lever for steering image-generating diffusion models in a 2025 paper. These models start from random noise and refine it into a picture. How the final image responds when the starting noise is tweaked was poorly understood, which limited control over sampling.
The team observed that the response is strikingly linear: scale the noise change, and the output changes in proportion. They back the observation with theory and experiments. On that basis they build CCS, a sampling method with a controller that nudges the starting noise to hit desired statistical traits. Tests showed tighter control than rival methods while keeping image quality and variety high.
Catalog record
| CREATOR | Bowen Song, Zecheng Zhang, Zhaoxu Luo, Jason Hu, Wei Yuan, Jing Jia, Zhengxu Tang, Guanyang Wang, Liyue Shen |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 58 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-04670