Generative diffusion models from a PDE perspective
Fei Cao, Kimball Johnston, Thomas Laurent, Justin Le, Sébastien Motsch · 2025
Read Generative diffusion models from a PDE perspective by Fei Cao, Kimball Johnston, Thomas Laurent, Justin Le, Sébastien Motsch (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-2501-17054.txt.
Description
Fei Cao and co-authors explain AI image-making diffusion models through the lens of equations for change over time. These models learn by slowly adding noise to data, then learning to undo that noise step by step. Their 2025 paper asks what the math of the reverse path really shows.
Working through the equations, the team show the reverse path stays inside the spread of the source data. In pure form, the method adds no smoothing of its own, so new-sample skill must come from elsewhere. They also solve the reverse path when the start point is fixed. For a finite set of points, exact solving just returns the old points, a sign of overfitting.
Catalog record
| CREATOR | Fei Cao, Kimball Johnston, Thomas Laurent, Justin Le, Sébastien Motsch |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 64 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-17054