Infinite-dimensional generative diffusions via Doob’s h-transform
Thorben Pieper-Sethmacher, Daniel Paulin · 2026
Read Infinite-dimensional generative diffusions via Doob’s h-transform by Thorben Pieper-Sethmacher, Daniel Paulin (2026) 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-2602-06621.txt.
Description
Thorben Pieper-Sethmacher and Daniel Paulin give a careful mathematical treatment of generative diffusion models, the tools behind much modern AI data creation. Most such models are built by reversing a process that adds noise. The authors instead start from a reference process and steer it toward the target data by reweighting its paths.
This steering trick, based on a classic change of measure, extends cleanly to infinite-dimensional settings where the usual reversal approach is hard to define. The paper proves the construction works under checkable conditions and bounds how close the result sits to the target. It also shows the steered process can be learned through a score-matching objective, and tests the idea on both synthetic and real data.
Catalog record
| CREATOR | Thorben Pieper-Sethmacher, Daniel Paulin |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 65 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-06621