Overcoming the Curvature Bottleneck in MeanFlow
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2025
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866911551868370944 |
|---|---|
| author | Zhang, Xinxi Tan, Shiwei Nguyen, Quang Dao, Quan Han, Ligong He, Xiaoxiao Zhang, Tunyu Mao, Chengzhi Metaxas, Dimitris Pavlovic, Vladimir |
| author_facet | Zhang, Xinxi Tan, Shiwei Nguyen, Quang Dao, Quan Han, Ligong He, Xiaoxiao Zhang, Tunyu Mao, Chengzhi Metaxas, Dimitris Pavlovic, Vladimir |
| contents | MeanFlow offers a promising framework for one-step generative modeling by directly learning a mean-velocity field, bypassing expensive numerical integration. However, we find that the highly curved generative trajectories of existing models induce a noisy loss landscape, severely bottlenecking convergence and model quality. We leverage a fundamental geometric principle to overcome this: mean-velocity estimation is drastically simpler along straight paths. Building on this insight, we propose Rectified MeanFlow, a self-distillation approach that learns the mean-velocity field over a straightened velocity field, induced by rectified couplings from a pretrained model. To further promote linearity, we introduce a distance-based truncation heuristic that prunes residual high-curvature pairs. By smoothing the optimization landscape, our method achieves strong one-step generation performance. We improve the FID of baseline MeanFlow models from 30.9 to 8.6 under same training budget, and outperform the recent 2-rectified flow++ by 33.4% in FID while running 26x faster. Our work suggests that the difficulty of one-step flow generation stems partially from the rugged optimization landscapes induced by curved trajectories. Code is available at https://github.com/Xinxi-Zhang/Re-MeanFlow. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2511_23342 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Overcoming the Curvature Bottleneck in MeanFlow Zhang, Xinxi Tan, Shiwei Nguyen, Quang Dao, Quan Han, Ligong He, Xiaoxiao Zhang, Tunyu Mao, Chengzhi Metaxas, Dimitris Pavlovic, Vladimir Computer Vision and Pattern Recognition Artificial Intelligence MeanFlow offers a promising framework for one-step generative modeling by directly learning a mean-velocity field, bypassing expensive numerical integration. However, we find that the highly curved generative trajectories of existing models induce a noisy loss landscape, severely bottlenecking convergence and model quality. We leverage a fundamental geometric principle to overcome this: mean-velocity estimation is drastically simpler along straight paths. Building on this insight, we propose Rectified MeanFlow, a self-distillation approach that learns the mean-velocity field over a straightened velocity field, induced by rectified couplings from a pretrained model. To further promote linearity, we introduce a distance-based truncation heuristic that prunes residual high-curvature pairs. By smoothing the optimization landscape, our method achieves strong one-step generation performance. We improve the FID of baseline MeanFlow models from 30.9 to 8.6 under same training budget, and outperform the recent 2-rectified flow++ by 33.4% in FID while running 26x faster. Our work suggests that the difficulty of one-step flow generation stems partially from the rugged optimization landscapes induced by curved trajectories. Code is available at https://github.com/Xinxi-Zhang/Re-MeanFlow. |
| title | Overcoming the Curvature Bottleneck in MeanFlow |
| topic | Computer Vision and Pattern Recognition Artificial Intelligence |
| url | https://arxiv.org/abs/2511.23342 |