Current diffusion pipelines still require dozens of denoising steps to reach ImageNet‑level fidelity, and the compute cost scales linearly with step count. Representation Distribution Matching flips that assumption by delivering ImageNet‑quality samples in a single forward pass, dramatically reducing the inference cost compared to multi‑step methods [1].
Until now, high‑resolution synthesis has been dominated by multi‑step samplers such as DDPM, PLMS, and recent four‑step variants of FLUX.2, all of which trade speed for quality measured by FID or similar perceptual metrics. Practitioners have accepted the latency as inevitable because no single‑step method could match the visual fidelity of these multi‑step baselines.
iRDM sets a new one‑step state of the art on ImageNet with an SW_r14 score of 1.30, eclipsing every prior single‑step generator and confirming its superiority via PickScore, which prefers it over the former best in 71.2 % of paired samples [1].
The same recipe also rewrites a four‑step FLUX.2 model into a one‑step generator that outperforms its multi‑step teacher on GenEval (0.826 versus 0.794) and PickScore (22.76 versus 22.58), all while consuming just 90 H200 GPU‑hours of training [1].
The method’s gains hinge on two fragile design choices: batch sizes must exceed 2,048 to stabilize the MMD estimate, and a balanced battery of frozen encoders is required because any single representation can be gamed into deceptively low scores while images remain visibly fake. These constraints leave open whether the approach scales to non‑ImageNet domains or smaller compute budgets.
If one‑step diffusion truly matches multi‑step quality, the community should replace FID‑oriented baselines with multi‑encoder SW_r14 reporting and treat iRDM as the default efficiency target for large‑scale generative deployments.
Top comments (0)