Consistent Distribution Matching for Data-Free Diffusion Distillation

2026-10-06

AI Summary

This study addresses the slow inference of diffusion models and the reliance of existing distillation methods on training data or complex architectures by proposing a data-free, simulation-free distillation framework. Methodologically, it unifies generation and score estimation within a single network and simplifies training by optimizing only a single objective between the teacher and student models. Theoretically, this work provides the first proof that minimizing this objective guarantees Wasserstein convergence of the student flow map toward the teacher distribution. Empirically, the proposed approach achieves an FID of 2.04 with one neural function evaluation (1-NFE) and 1.37 with 4-NFE on ImageNet, significantly outperforming existing state-of-the-art data-free baselines.

Abstract

Flow and diffusion models suffer from slow inference due to computationally expensive numerical integration. Distillation provides a promising way for a student model to learn from a teacher's dynamics, enabling one-step or few-step generation. However, existing methods often depend on curated distillation datasets, costly teacher rollouts, or auxiliary proxy networks, which complicate model training and scaling. In this work, we propose Consistent Distribution Matching, a simulation-free and data-free distillation method for accelerating diffusion and flow models while preserving strong generative capacity. Our key insight is to unify sample generation and score estimation with one student network. Thus, our framework uses only two models, a frozen teacher and a trainable student, and optimizes one objective. We prove that minimizing our objective indicates Wasserstein convergence of the student flow-map pushforwards to the teacher marginals. On ImageNet 256$\times$256, our method attains an FID of 2.04 with a single function evaluation (1-NFE) and a 4-NFE FID of 1.37 within 40 epochs of training, surpassing the state-of-the-art distillation baselines without data. Our code code and model are available at https://consistentdmd.github.io/.

Problem

• Diffusion Distillation
• Data-Free
• Flow Models
• Inference Acceleration

Innovation

• Data-Free Distillation
• Consistent Distribution Matching
• Diffusion Models
• Wasserstein Convergence
• Flow Matching