Representation Distribution Matching for One-Step Visual Generation

2026-07-02 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, quick

Summary

The Representation Distribution Matching (RDM) paradigm trains one-step image generators by aligning generated and reference feature distributions using frozen pretrained encoders. Researchers identified two design axes: how distributions are compared and the representations used. Key findings include that the classical MMD objective, when estimated correctly, is effective and scalable, and an optimal generated batch size exceeds 2048. Furthermore, relying on a single representation can be gamed, leading to visibly fake images despite good scores; thus, matching against a balanced battery of encoders and evaluating with SW_r14, a Sliced-Wasserstein distance over 14 encoders, is crucial. The improved RDM (iRDM) achieves a one-step state-of-the-art on ImageNet with SW_r14 1.30, preferred over the prior best one-step generator on 71.2% of samples by PickScore. iRDM also post-trains the four-step FLUX.2 into a one-step generator, surpassing its four-step version on GenEval (0.826 to 0.794) and PickScore (22.76 to 22.58) in 90 H200 GPU-hours.

Key takeaway

For Machine Learning Engineers developing efficient visual generation models, you should reconsider classical MMD objectives, ensuring proper estimation and utilizing larger batch sizes, potentially exceeding 2048. To ensure robust model evaluation and prevent gaming, integrate a balanced battery of encoders, such as the SW_r14 metric. This approach can significantly improve one-step generator performance and efficiently convert multi-step models like FLUX.2 into faster, high-performing single-step versions.

Key insights

Correctly estimating MMD and using diverse encoders enables robust one-step visual generation.

Principles

• MMD is effective for distribution matching when estimated right.
• Optimal batch sizes for generation can exceed 2048.
• Evaluate with multiple encoders to prevent gaming single metrics.

Method

Train a one-step generator by matching generated and reference feature distributions using frozen pretrained encoders, optimizing MMD with large batch sizes and evaluating with SW_r14.

In practice

• Apply iRDM to achieve state-of-the-art one-step ImageNet generation.
• Post-train multi-step models like FLUX.2 into efficient one-step versions.

Topics

• Representation Distribution Matching
• One-Step Visual Generation
• ImageNet
• MMD Objective
• FLUX.2
• Generative Models
• Feature Distribution Matching

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• More Than Image Generators: A Science of Problem-Solving using Probability | Diffusion Models — Diffusion models combine deep learning with Langevin sampling to generate images by iteratively following noisy gradients of an image's probability distribution.
• Weighted quantization using MMD: From mean field to mean shift via gradient flows — MMD minimization via WFR geometry and MSIP offers robust, efficient probability distribution quantization with variable particle weights.
• MilliVid: Hierarchical Latents for Long-Range Consistency in Video Generation — Hierarchical latent token compression and coarse-to-fine rollout enable long-range consistency in video generation by managing detail levels efficiently.
• High-Fidelity Two-Step Image Generation via Teacher-Aligned End-to-End Distillation — Tailored distillation strategies can significantly improve the quality-efficiency trade-off for few-step image generation.
• MaSC: A Masked Similarity Metric for Evaluating Concept-Driven Generation — Spatially decomposed aggregation improves evaluation of concept-driven image generation.
• Calibrating Generative Models to Feature Distributions with MMD Finetuning — kCGM corrects generative model miscalibration by matching feature distributions with MMD and KL regularization, improving validity.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.