Continuous-Time Distribution Matching for Few-Step Diffusion Distillation

2026-05-07

AI Summary

This work proposes a continuous-time distribution matching distillation (DMD) framework that overcomes key limitations of conventional DMD, which relies on discrete timesteps and often suffers from visual artifacts and oversmoothing, necessitating auxiliary modules such as GANs for image quality restoration. By introducing a dynamic continuous scheduling scheme and a novel trajectory alignment objective based on extrapolation from the student’s velocity field, the method enables distribution matching at arbitrary sampling points without requiring complex post-hoc refinements. Combined with reverse KL divergence optimization and a stochastic training length strategy, the approach significantly enhances few-step generation fidelity on models such as SD3-Medium and Longcat-Image, achieving visual quality on par with state-of-the-art methods.

Abstract

Step distillation has become a leading technique for accelerating diffusion models, among which Distribution Matching Distillation (DMD) and Consistency Distillation are two representative paradigms. While consistency methods enforce self-consistency along the full PF-ODE trajectory to steer it toward the clean data manifold, vanilla DMD relies on sparse supervision at a few predefined discrete timesteps. This restricted discrete-time formulation and mode-seeking nature of the reverse KL divergence tends to exhibit visual artifacts and over-smoothed outputs, often necessitating complex auxiliary modules -- such as GANs or reward models -- to restore visual fidelity. In this work, we introduce Continuous-Time Distribution Matching (CDM), migrating the DMD framework from discrete anchoring to continuous optimization for the first time. CDM achieves this through two continuous-time designs. First, we replace the fixed discrete schedule with a dynamic continuous schedule of random length, so that distribution matching is enforced at arbitrary points along sampling trajectories rather than only at a few fixed anchors. Second, we propose a continuous-time alignment objective that performs active off-trajectory matching on latents extrapolated via the student's velocity field, improving generalization and preserving fine visual details. Extensive experiments on different architectures, including SD3-Medium and Longcat-Image, demonstrate that CDM provides highly competitive visual fidelity for few-step image generation without relying on complex auxiliary objectives. Code is available at https://github.com/byliutao/cdm.

Problem

• Distribution Matching Distillation
• diffusion distillation
• visual artifacts
• over-smoothing
• few-step generation

Innovation

• Continuous-Time Distribution Matching
• Diffusion Distillation
• Few-Step Generation
• Trajectory Alignment
• Velocity Field Extrapolation