RDM: Representation Distribution Matching for One-Step Visual Generation

Model Overview

Representation Distribution Matching (RDM / iRDM) is a paradigm for training state-of-the-art one-step visual generation models developed by researchers from EPFL (VITA Lab), Valeo.ai, and Sorbonne Université (Lanfeng Judge, et al.). It matches feature distributions between generated and reference images using Maximum Mean Discrepancy (MMD) with Nyström estimation across a battery of frozen pretrained encoders.

Post-training multi-step diffusion models (such as FLUX.2 Klein) with RDM distills them into single-step generators that match or exceed the visual quality and prompt adherence of their multi-step teacher models.

---

Key Features

• Representation Distribution Matching (RDM): Directly matches generated image distributions with real image feature distributions using MMD across frozen pretrained visual encoders.
• Improved RDM (iRDM): Combines within-batch repulsion with Nyström landmark estimation for efficient feature attraction.
• Multi-Encoder Battery Defense: Evaluated across a battery of 14 frozen encoders (SW_r14 metric) to prevent model representation gaming.
• Single-Step FLUX.2 Distillation: Post-trains FLUX.2 Klein into a 1-step generator outperforming its 4-step teacher on key visual benchmarks.
• Large-Batch MMD Estimation: Utilizes large batch training to ensure stable gradient estimation for single-step generation.

---

Verified Project Links

• Project Website: https://alan-lanfeng.github.io/rdm/
• arXiv Paper: https://arxiv.org/abs/2607.02375
• GitHub Repository: https://github.com/vita-epfl/RDM
• Hugging Face Model: https://huggingface.co/epfl-vita/flux2-klein-1step-rdm

---

Benchmarks & Results

• GenEval (FLUX.2 Klein 1-Step): 0.826 score (vs. 0.794 for the 4-step teacher model).
• PickScore (FLUX.2 Klein 1-Step): 22.76 (vs. 22.58 for teacher model).
• Human Preference: 71.2% preference win-rate over prior single-step diffusion baselines.