DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models
DiffusionART is presented as a reinforcement method that makes diffusion fine-tuning faster and more stable without EMA tuning.
The paper says existing reverse-process RL methods suffer from weak credit assignment, while NFT-style forward methods can be efficient but unstable. Its method aggregates rewards per prompt into a shared reward-improving direction, then adds a trust-region-masked restoration update. The authors report that on HunyuanVideo-13B, it reached comparable late-stage reward levels with 12.02x and 11.72x fewer GPU hours than DiffusionNFT and Flow-GRPO-Fast. It also supports pairwise reward models and does not require a differentiable reward model. HF Daily Papers' note
The paper says existing reverse-process RL methods suffer from weak credit assignment, while NFT-style forward methods can be efficient but unstable. Its method aggregates rewards per prompt into a shared reward-improving direction, then adds a trust-region-masked restoration update. The authors report that on HunyuanVideo-13B, it reached comparable late-stage reward levels with 12.02x and 11.72x fewer GPU hours than DiffusionNFT and Flow-GRPO-Fast. It also supports pairwise reward models and does not require a differentiable reward model. HF Daily Papers' note
score 5