Consistent Distribution Matching for Data-Free Diffusion Distillation
Abstract
Consistent Distribution Matching accelerates diffusion and flow models via simulation-free, data-free distillation that unifies generation and score estimation, achieving strong one-step and few-step results.
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 256times256, 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/.
Community
Excited to share our work on Consistent Distribution Matching! 🚀
Inspired by Distribution Matching Distillation (DMD) and Score Identity Distillation (SiD), we ask a simple question: Can we achieve distribution matching without training an auxiliary fake-score network?
Our answer is yes!
CDMD takes a step further by unifying sample generation and velocity estimation within a single student network. This gives us a much simpler distillation framework:
One frozen teacher + one trainable student + one objective.
No external training data, teacher trajectory simulation, or auxiliary score networks in our data-free setting.
Stronger generative performance: 2.04 FID at 1 NFE on ImageNet 256×256, outperforming evaluated data-free baselines (DMD2: 3.43 FID).
Near-teacher quality: 1.37 FID at 4 NFEs, approaching the teacher's 1.35 FID.
We also provide theoretical insights connecting our consistency objective to distribution matching through Wasserstein convergence.
Our takeaway: Effective distribution matching doesn't have to come with complicated multi-network optimization.
We'd love to hear your thoughts and feedback!
Get this paper in your agent:
hf papers read 2610.09221 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 1
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper