A visual-generation trainer makes reward learning faster and steadier

Can reward-guided training improve the visual qualities people care about without forcing each new model or task to inherit a fragile stabilization schedule?

The paper examines whether instability comes from how reward information is assigned across denoising states, and whether a method can preserve useful guidance while restoring overly large updates.

DiffusionART turns each prompt’s rewarded outputs into a shared update, then restores drifted states so visual generators learn rewards efficiently without EMA tuning.

The authors show that NFT’s two reward branches share useful guidance but retain a state-dependent consistency force.

In controls, removing the old-policy anchor still collapses; two updates without EMA avoid collapse over the observed window but take 15.2 times longer, pointing to restoration as the stabilizer.

DiffusionART applies a shared reward-endpoint direction to visited states, then restores drift. In HunyuanVideo-13B, it reaches VideoReward’s visual-quality score of 5.5 in 24 steps versus 273 for DiffusionNFT with EMA, then records higher visual- and motion-quality scores under the reported setup.

The speedups and evaluation scores come from the reported models, modalities, reward settings, sampling procedures, and training or evaluation windows. They do not establish the same results outside those tested configurations.