On the Stability and Convergence of Flow Map Training
Abstract
Flow maps are a recent method of probabilistic generation that amortize diffusion trajectories into one-step samplers, unlocking efficient generation. While progress has been made in improving their empirical performance, there is still a lack of understanding behind the mechanisms underlying their success. Specifically, many flow map learning methods employ stop gradient operators as a crucial part of their training target. Adaptive weighting also drastically improves the generation capabilities of flow maps. Although many interpretations have been proposed to support the inclusion of these heuristics, none have analyzed the impact of these assumptions on the training dynamics of flow maps themselves. In this work, we address this issue by providing grounded theory explaining why such tricks are required. Analyzing the functional gradient flow induced by the MeanFlow and Lagrangian losses, we show their convergence to the true flow map, settling the validity of these objectives. For finite updates, we highlight a severe roadblock to the usage of these losses in practice: the possible unboundedness of the update. Adaptive normalization not only provides bounded updates, but, combined with a specific weighting, additionally provides convergence guarantees towards a neighborhood of the solution. On synthetic data, we show that our functional gradient flow analysis exactly describes the true training dynamics for flow maps. Experiments on standard image datasets demonstrate that a proper reweighting scheme is essential for strong empirical performance, which is only achieved for a specific range of reweighting strengths.
Related
- Tangential stability and fully discrete convergence of the classical BGN scheme for curve shortening flow
- Quasi-convergence of stability conditions
- Convergence of a vector-valued Allen-Cahn system to Brakke's multiphase mean curvature flow
- Convergence of Generalized Kähler--Ricci Flow on Toric Fano Manifolds
- Convergence of the Calabi flow on Kähler surfaces with bounded Sobolev constants
BibTeX
@misc{nath2026stability,
title = {{On the Stability and Convergence of Flow Map Training}},
author = {Ishan Nath and Tom Stanic and Avishek Joey Bose and Gauthier Gidel and Quentin Bertrand},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.20674},
eprint = {2610.20674},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.20674},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - On the Stability and Convergence of Flow Map Training
AU - Ishan Nath
AU - Tom Stanic
AU - Avishek Joey Bose
AU - Gauthier Gidel
AU - Quentin Bertrand
PY - 2026
DA - 2026/10/14
PB - arXiv
DO - 10.48550/arxiv.2610.20674
UR - https://arxiv.org/abs/2610.20674
AB - Flow maps are a recent method of probabilistic generation that amortize diffusion trajectories into one-step samplers, unlocking efficient generation. While progress has been made in improving their empirical performance, there is still a lack of understanding behind the mechanisms underlying their success. Specifically, many flow map learning methods employ stop gradient operators as a crucial part of their training target. Adaptive weighting also drastically improves the generation capabilities of flow maps. Although many interpretations have been proposed to support the inclusion of these heuristics, none have analyzed the impact of these assumptions on the training dynamics of flow maps themselves. In this work, we address this issue by providing grounded theory explaining why such tricks are required. Analyzing the functional gradient flow induced by the MeanFlow and Lagrangian losses, we show their convergence to the true flow map, settling the validity of these objectives. For finite updates, we highlight a severe roadblock to the usage of these losses in practice: the possible unboundedness of the update. Adaptive normalization not only provides bounded updates, but, combined with a specific weighting, additionally provides convergence guarantees towards a neighborhood of the solution. On synthetic data, we show that our functional gradient flow analysis exactly describes the true training dynamics for flow maps. Experiments on standard image datasets demonstrate that a proper reweighting scheme is essential for strong empirical performance, which is only achieved for a specific range of reweighting strengths.
ER -