ProReflow: Progressive Reflow with Decomposed Velocity

2025-03-05

AI Summary

Diffusion models suffer from high computational costs during sampling. Method: This paper identifies suboptimality in the original flow matching training procedure and proposes Progressive Reflow—a framework that decomposes complex diffusion trajectories into local time intervals, enabling ultra-low-step generation via staged calibration. It introduces a novel direction-aligned v-prediction strategy, prioritizing flow direction matching over magnitude matching to significantly reduce training difficulty. Contribution/Results: On Stable Diffusion v1.5, Progressive Reflow achieves FID=10.70 on MSCOCO with only four sampling steps—approaching the performance of a 32-step DDIM teacher model (FID=10.05). To our knowledge, this is the first work to integrate progressive reflow with direction-aligned v-prediction, establishing a new paradigm for efficient, high-fidelity few-step generative modeling.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Search and Optimization: Sampling/Simulation-based Search
• Machine Learning: Deep Neural Architectures and Foundation Models

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Diffusion models have achieved significant progress in both image and video generation while still suffering from huge computation costs. As an effective solution, flow matching aims to reflow the diffusion process of diffusion models into a straight line for a few-step and even one-step generation. However, in this paper, we suggest that the original training pipeline of flow matching is not optimal and introduce two techniques to improve it. Firstly, we introduce progressive reflow, which progressively reflows the diffusion models in local timesteps until the whole diffusion progresses, reducing the difficulty of flow matching. Second, we introduce aligned v-prediction, which highlights the importance of direction matching in flow matching over magnitude matching. Experimental results on SDv1.5 and SDXL demonstrate the effectiveness of our method, for example, conducting on SDv1.5 achieves an FID of 10.70 on MSCOCO2014 validation set with only 4 sampling steps, close to our teacher model (32 DDIM steps, FID = 10.05).

Problem

• Reduces computation costs in diffusion models
• Improves flow matching training pipeline
• Enhances image generation with fewer steps

Innovation

• Progressive reflow reduces flow matching difficulty.
• Aligned v-prediction emphasizes direction matching importance.
• Achieves high FID with minimal sampling steps.