---
title: FLUX.2-klein 1-step RDM
emoji: ⚡
colorFrom: purple
colorTo: pink
sdk: gradio
sdk_version: 5.50.0
app_file: app.py
pinned: true
python_version: 3.11
license: apache-2.0
short_description: 1-step FLUX.2-klein that beats its 4-step teacher
models:
  - epfl-vita/flux2-klein-1step-rdm
  - black-forest-labs/FLUX.2-klein-4B
---

# FLUX.2-klein · 1-step · RDM (EPFL VITA)

Live demo of [`epfl-vita/flux2-klein-1step-rdm`](https://huggingface.co/epfl-vita/flux2-klein-1step-rdm) —
FLUX.2-klein-4B distilled from 4 denoising steps to **one** with Representation Distribution Matching (RDM).

| | GenEval ↑ | PickScore-COCO ↑ | steps |
|---|---|---|---|
| **this model** | **0.826** | **22.76** | **1** |
| FLUX.2-klein-4B teacher | 0.794 | 22.58 | 4 |

Pipeline: prompt → Qwen3-4B ctx → fresh noise → 1-step flow-matching Euler → FLUX.2 VAE decode
(identical to the evaluated pipeline). Runs on ZeroGPU; the first generation loads the models (~2 min).

## Paper

**Representation Distribution Matching for One-Step Visual Generation.**
[arXiv:2607.02375](https://arxiv.org/abs/2607.02375) · [Hugging Face Papers](https://huggingface.co/papers/2607.02375) · [Project page](https://alan-lanfeng.github.io/rdm/) · [Code](https://github.com/vita-epfl/RDM)

```bibtex
@article{feng2026rdm,
  title={Representation Distribution Matching for One-Step Visual Generation},
  author={Feng, Lan and Li, Wuyang and Zablocki, {\'E}loi and Cord, Matthieu and Alahi, Alexandre},
  journal={arXiv preprint arXiv:2607.02375},
  year={2026}
}
```