Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation

2026-10-02

AI Summary

Flow matching models incur substantial inference costs, and existing acceleration methods typically require retraining or compromise generation quality. This work proposes COFLOW, a plug-and-play framework that adaptively selects the number of generation steps based on prompt features, enabling efficient visual generation without retraining. The method dynamically optimizes the sampling strategy by integrating forward Euler discretization error analysis with unsupervised reward-based online reinforcement learning. Experiments demonstrate that COFLOW achieves a 2.5× speedup in both image and video generation tasks while preserving perceptual and semantic quality. Furthermore, this study provides rigorous theoretical error bounds for the proposed approach.

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.

Problem

• Flow Matching
• Visual Generation
• Inference Acceleration
• Adaptive Step Selection

Innovation

• Flow Matching
• Adaptive Step Selection
• Contextual Inference
• Plug-and-Play Acceleration
• Visual Generation