Learning Straight Flows by Learning Curved Interpolants
Shiv Shankar, Tomas Geffner · 2025
Read Learning Straight Flows by Learning Curved Interpolants by Shiv Shankar, Tomas Geffner (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-2503-20719.txt.
Description
Shiv Shankar and Tomas Geffner proposed learning straighter generative flows in a 2025 paper. Flow matching models ferry noise into data along paths set by interpolants, usually straight lines. Paired with independent noise-data matching, straight interpolants still yield a curved velocity field, which forces slow, many-step generation. Their idea flips the problem: bend the interpolant so the learned field comes out straight.
They set the search up as a multi-level optimization and give an efficient approximate solver. The procedure trains end to end without simulation. Straighter fields mean fewer steps at generation time and faster sampling. The framework is general: any flow model stuck with curved trajectories can, in principle, trade interpolant shape for inference speed, the authors argue.
Catalog record
| CREATOR | Shiv Shankar, Tomas Geffner |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 28 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-20719