Unifying Distributional Training for One-Step Visual Generation

TL;DR

The paper presents MGFlow, a unified framework for one-step visual generation that uses Gaussian mixture models to represent feature distributions and combines optimal transport with score-based matching. It improves upon existing methods by addressing mode collapse through mass-constrained sample assignment and component-level updates, achieving state-of-the-art performance on ImageNet and enabling high-quality text-to-image generation from a pre-trained diffusion model.

Why It Matters

One-step visual generation is critical for real-time applications, and this work advances the field by providing a theoretically grounded, scalable method that outperforms multi-step models in quality and efficiency. The ability to post-train existing diffusion models into one-step generators offers practical benefits for deployment in resource-constrained environments.

Key Findings

• MGFlow introduces a unified framework that connects global objectives to pointwise updates via Wasserstein gradient flow, enabling flexible distributional training.
• The method uses Gaussian mixture models with adjustable granularity to balance global moments and sample-level representation, improving expressivity and stability.
• By integrating mass-constrained sample assignment and paired component updates, MGFlow effectively mitigates mode collapse beyond what mixture modeling alone can achieve.