Noise is All You Need: Solving Linear Inverse Problems by Noise Combination Sampling with Diffusion Models
Xun Su Hiroyuki Kasai · 2025
Read Noise is All You Need: Solving Linear Inverse Problems by Noise Combination Sampling with Diffusion Models by Xun Su Hiroyuki Kasai (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-2510-23633.txt.
Description
Xun Su and a co-author present a new way to solve linear inverse problems — recovering a clean signal from blurred or partial measurements — using pretrained diffusion models. These models can fill in missing content, but steering them with the measurements is a balancing act. Push the data in too hard and the generation breaks; push too softly and the answer ignores the measurements.
Their method, Noise Combination Sampling, builds the best noise vector from a whole noise subspace, chosen to mimic the measurement score. That crafted noise replaces the usual noise term in the diffusion process, so the conditions enter naturally, with no step-by-step tuning knobs. The method plugs into many existing solvers, even image compression, and shines when few generation steps are allowed. Tests show better results with almost no extra computing cost, plus improved stability.
Catalog record
| CREATOR | Xun Su Hiroyuki Kasai |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 48 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2510-23633