Overcoming the Curvature Bottleneck in MeanFlow
arXiv:2511.23342v3 Announce Type: replace-cross Abstract: 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 appro — Xinxi Zhang, Shiwei Tan, Quang Nguyen, Quan Dao, Ligong Han, Xiaoxiao He, Tunyu Zhang, Chengzhi Mao, Dimitris Metaxas, Vladimir Pavlovic
View PDF HTML (experimental)
Abstract: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 this https URL.
Subjects:
Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
Cite as: arXiv:2511.23342 [cs.CV]
(or arXiv:2511.23342v3 [cs.CV] for this version)
https://doi.org/10.48550/arXiv.2511.23342
arXiv-issued DOI via DataCite
Submission history
From: Xinxi Zhang [view email] [v1] Fri, 28 Nov 2025 16:50:08 UTC (25,518 KB) [v2] Thu, 12 Mar 2026 20:20:16 UTC (22,478 KB) [v3] Sun, 29 Mar 2026 19:55:09 UTC (25,902 KB)
Sign in to highlight and annotate this article
Conversation starters
Daily AI Digest
Get the top 5 AI stories delivered to your inbox every morning.
More about
researchpaperarxiv
How to Build an AI Content Playbook That Actually Protects Your Voice
Ahnii! You've read the articles warning you not to let AI take over your content. Ruth Doherty's latest piece is one of the best: a clear-eyed breakdown of where AI helps and where it silently destroys your brand. This post shows you how to take that framework and turn it into an actual operating document for your content pipeline. Why a Framework Without a Playbook Doesn't Stick Ruth's core argument is sharp: AI is an efficiency engine, not a strategy engine. Use it for research, structuring, repurposing, and editing. Keep it away from messaging, customer research, and anything that requires your actual point of view. That distinction is easy to agree with. It's harder to enforce on a Tuesday afternoon when you're behind on three social posts and the AI can draft all of them in 90 seconds
Top 10 Best Universities to Study AI in USA 2026 Led by CMU and MIT With Strong Research and Industry Ties - International Business Times Australia
Top 10 Best Universities to Study AI in USA 2026 Led by CMU and MIT With Strong Research and Industry Ties International Business Times Australia
Knowledge Map
Connected Articles — Knowledge Graph
This article is connected to other articles through shared AI topics and tags.
Discussion
Sign in to join the discussion
No comments yet — be the first to share your thoughts!