Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation
COFLOW chooses generation step counts at inference time from the prompt, aiming to cut compute without retraining the model.
The paper says the method trains online with an unsupervised reward that trades off speed and fidelity. It is described as plug-and-play for existing flow models and tested on both image and video generation. The authors report more than a 2.5x speedup while preserving perceptual and semantic quality, and include an O(1/K) forward-Euler error bound under standard assumptions. ArXiv · AI/CL/LG's note
The paper says the method trains online with an unsupervised reward that trades off speed and fidelity. It is described as plug-and-play for existing flow models and tested on both image and video generation. The authors report more than a 2.5x speedup while preserving perceptual and semantic quality, and include an O(1/K) forward-Euler error bound under standard assumptions. ArXiv · AI/CL/LG's note
score 5