Rethinking Diffusion Posterior Sampling: From Conditional Score Estimator to Maximizing a Posterior
Tongda Xu, Xiyan Cai, Xinjie Zhang, Xingtong Ge, Dailan He, Ming Sun, Jingjing Liu, Ya-Qin Zhang, Jian Li, Yan Wang · 2025
Read Rethinking Diffusion Posterior Sampling: From Conditional Score Estimator to Maximizing a Posterior by Tongda Xu, Xiyan Cai, Xinjie Zhang, Xingtong Ge, Dailan He, Ming Sun, Jingjing Liu, Ya-Qin Zhang, Jian Li, Yan Wang (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-2501-18913.txt.
Description
Tongda Xu and a large team rethink a popular diffusion based image reconstruction method in a 2025 paper. Diffusion posterior sampling is widely used to solve inverse problems without extra training. It was believed to work by approximating a conditional score. Testing on large images, the team show that approximation actually strays far from a true conditional score. It is even worse than the unconditional score, and its average is far from zero, which a valid score should not be.
They also find the method's samples are high quality but much less diverse than true posterior samples. Together, the evidence says the method behaves like maximizing a posterior instead. The team then lean into that reading: they maximize the posterior directly with gradient steps and add a light conditional score estimator trained on few images. The improved version performed significantly better.
Catalog record
| CREATOR | Tongda Xu, Xiyan Cai, Xinjie Zhang, Xingtong Ge, Dailan He, Ming Sun, Jingjing Liu, Ya-Qin Zhang, Jian Li, Yan Wang |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 57 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-18913