A new way to generate AI images in just one step
Researchers just open-sourced RDM, a technique that generates high-quality images in a single forward pass instead of multi-step diffusion. By matching visual feature distributions across frozen encoders, it compressed a 4-step FLUX.2 into a 1-step model that actually beats its teacher. It cuts generation latency to virtually instant while keeping compute requirements modest.
RDM (Representation Distribution Matching) is a new open-source recipe that condenses image generation down to a single forward pass.
Most modern diffusion models still take 30 to 50 denoising steps to render a clean image. Even modern fast variants like Turbo or Flash models demand 4 to 10 steps. That still means multiple rounds of compute, which stalls real-time apps. RDM aims to wipe out that latency entirely with just one evaluation.
How it hits one step
Traditional distillation methods usually lean on heavy online teachers, tricky adversarial losses, or precise trajectory tracking. They are notoriously brittle to train and prone to mode collapse.
RDM skips all of that. There is no online teacher, no adversary, and no trajectory matching. Instead, it directly matches the probability distribution of generated features against real images across frozen, pretrained visual encoders.
The real trick lies in how those distributions are checked. Historically, matching features under a single encoder like DINOv2 leads to gaming: the network cheats the metric while the output still looks clearly fake. RDM prevents this by deploying a diverse battery of encoders held in balance by a Lagrangian controller. If the generator tries to cheat one space, the system automatically dials up the penalty there.
For text-to-image tasks, RDM matches the joint text-image distribution, forcing the model to honor complex prompts rather than just drifting toward generic pretty pictures.
Beating the four-step teacher
The numbers back up the theory. On the multi-encoder SW r14 distance metric, where real images sit at 1.00, RDM scores 1.30. That comfortably beats prior one-step state of the art (2.05) and even slips past several multi-step diffusion pipelines.
In human-preference testing via PickScore, a metric never seen during training, RDM samples were preferred over real photographs 63.6% of the time. It also beat the prior best one-step generator in 71.2% of head-to-head comparisons.
The most practical proof came from post-training FLUX.2. The team took the four-step FLUX.2 [klein] and converted it into a single-step generator in just 90 H200 GPU-hours. On the GenEval benchmark, the one-step model scored 0.826, topping the four-step teacher's 0.794. Its PickScore similarly rose from 22.58 to 22.76.
The code and local setup instructions are already published on the project page for anyone looking to reproduce the pipeline.
Why it matters
Cutting inference from four steps to one is a massive deal for real-time creativity. It opens the door for instant canvas painting, live UI skinning, and dramatically cheaper API inference at scale.
For a long time, the consensus was that single-step models were strictly for rough drafts. RDM shows that with better distribution matching and multi-encoder constraints, one-step generators can actually outshoot the multi-step models they were built from.
Minor artifacts can still crop up on intricate details given the single-shot constraint. But if you are building real-time visual features or trying to slash GPU serving bills, this repository belongs in your testing queue right now.
Sources (1)
- alan-lanfeng.github.io(opens in a new tab)Originalalan-lanfeng.github.io