arXiv：cs.AI· Divya Jyoti Bajpai, Arun Verma, Manjesh Kumar Hanawal·· 4 小时前AI 评分40
COFLOW：面向高效视觉生成的 Flow Matching 自适应步数选择方法
Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation
AI 导读
COFLOW 是一种推理时方法，可根据提示词特征为每次生成自适应选择步数，无需重新训练底层生成模型。该方法通过无监督奖励在线训练，在图像和视频生成上实现超过 2.5 倍加速，同时保持感知与语义质量，并给出 O(1/K) 前向欧拉离散误差界。该工作已被 NeurIPS 2026 接收。
正文
Abstract:Flow Matching enables high-quality visual generation via continuous-time dynamics, but inference remains costly due to multiple sequential function evaluations. Existing acceleration methods reduce the number of function evaluations but often introduce additional training overhead, degrade quality, or fail to account for input-dependent variability. We propose COFLOW, an inference-time method that adaptively selects the step counts each generation based on the prompt features. Our context-aware COFLOW is trained online with an unsupervised reward that balances inference efficiency and generation fidelity. Our method is plug-and-play, requiring no retraining of the underlying generative model. It generalizes to image and video generation, achieving over 2.5x speedup while preserving perceptual and semantic quality. We further provide a theoretical analysis establishing an O(1/K) forward-Euler discretization error bound under standard regularity conditions.
| Comments: | Accepted in NeurIPS 2026 |
| Subjects: | Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.03202 [cs.CV] |
| (or arXiv:2610.03202v1 [cs.CV] for this version) | |
| https://doi.org/10.48550/arXiv.2610.03202 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Manjesh Kumar Hanawal [view email]
[v1]
Fri, 2 Oct 2026 12:19:49 UTC (5,297 KB)
来源：arXiv：cs.AI · arxiv.org