One-Step Diffusion with Distribution Matching Distillation

By Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Frédo Durand, William T. Freeman, Taesung Park. Computer Vision and Pattern Recognition, 2023

926 citations recorded by the source corpus.

Diffusion models generate high-quality images but require dozens of forward passes. We introduce Distribution Matching Distillation (DMD), a procedure to transform a diffusion model into a one-step image generator with minimal impact on image quality. We enforce the one-step image generator match the diffusion model at distribution level, by minimizing an approximate KL divergence whose gradient can be expressed as the difference between 2 score functions, one of the target distribution and the other of the synthetic distribution being produced by our one-step generator. The score functions are parameterized as two diffusion models trained separately on each distribution. Combined with a simple regression loss matching the large-scale structure of the multi-step diffusion outputs, our method outperforms all published few-step diffusion approaches, reaching 2.62 FID on ImageNet 64×64 and 11.49 FID on zero-shot COCO-30k, comparable to Stable Diffusion but orders of magnitude faster. Utilizing FP16 inference, our model can generate images at 20 FPS on modern hardware.

Subfields: Computational Mechanics, Control and Systems Engineering

View Paper: https://www.semanticscholar.org/paper/516a4c05883d4dc130bb7a638ad8e427b6756aac

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computational Mechanics

In Control and Systems Engineering

Citations in the corpus

Hunyuan-DiT: A Powerful Multi-Resolution Diffusion Transformer with Fine-Grained Chinese Understanding (2024, methodology): “Algorithmic Acceleration Recently, various methods have been proposed to reduce the inference steps of diffusion-based text-to-image models [22, 31, 21, 37, 40].”
One Step Diffusion via Shortcut Models (2024, neutral): “Alternatives to L2 distance have been proposed for distillation targets, such as adversarial (Sauer et al., 2023) or distribution-matching (Yin et al., 2024b;a) objectives.”
Improved Distribution Matching Distillation for Fast Image Synthesis (2024, methodology): “Recent works have extended score distillation [36, 37, 40–42] to diffusion distillation [22, 43–45].”
Fast High-Resolution Image Synthesis with Latent Adversarial Diffusion Distillation (2024, neutral): “Very recent distillation works aim at reducing the number of model evaluations to a single step, enabling real-time synthesis [63, 34, 62, 49, 28].”
Mean Flows for One-step Generative Modeling (2025, neutral): “One approach is to distill a pre-trained many-step diffusion model into a few-step model, e.g., [39, 13, 41] or score distillation [32, 50, 53].”
Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion (2025, neutral): “…measures and distribution matching frameworks, and we consider three approaches in this paper: • Distribution Matching Distillation (DMD) [98, 99]: This approach minimizes the reverse Kullback-Leibler divergence E t [ D KL ( p θ,t ∥ p data ,t )] by leveraging the score difference between…”
InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation (2023, methodology): “Inference Time FID-5k CLIP SD 1.4-DPM Solver (25 step)[5, 19] 0”
PixArt-Σ: Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation (2024, methodology): “Initially, we set T to the same value as the denoising timesteps during training, that is T = 999 following [52].”
From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (2024, neutral): “Distribution matching distillation is a technique designed to distill a slow, multi-step teacher diffusion model into an efficient few-step student model [99, 100].”