A research paper published on Hugging Face on October 8 and submitted on October 12, titled DiffusionART, proposes a new method for fine-tuning diffusion models with reinforcement learning. The authors, led by Zhenglin Wan and including researchers from KlingAI Research, say existing approaches have trade-offs: reverse-process methods such as Flow-GRPO are limited in efficiency by poor credit assignment, while forward-process methods such as DiffusionNFT are efficient but unstable without carefully tuned exponential-moving-average (EMA) training of off-policy samples.

The paper says NFT-style objectives contain a hidden forward-reverse consistency force that cancels only partially when positive and negative training are combined, and that EMA stabilizes training through restoration rather than off-policy rollouts. To address this, the authors introduce DiffusionART (Diffusion Reinforcement with Aggregation-First Reward Tilting). It aggregates each prompt's rewards into a shared reward-improving direction for reverse-process states, removes the consistency force by construction, and adds a trust-region-masked second update for drift-adaptive restoration without EMA tuning.

According to the paper, DiffusionART requires no differentiable reward model, remains effective with fewer samples per prompt, and supports pairwise reward models (DiffusionART-Pairwise). Across image and video generation, it trains efficiently and stably without per-task tuning. On HunyuanVideo-13B with equally weighted visual and motion quality rewards, the authors report it reaches the late-stage mean reward levels of DiffusionNFT and Flow-GRPO-Fast with 12.0x and 11.7x fewer GPU hours, respectively, while achieving 28.8% and 53.0% higher mean training rewards. The paper was ranked #1 Paper of the Day on Hugging Face.

A research paper published on Hugging Face on October 8 and submitted on October 12, titled DiffusionART, proposes a new method for fine-tuning diffusion models with reinforcement learning. The authors, led by Zhenglin Wan and including researchers from KlingAI Research, say existing approaches have trade-offs: reverse-process methods such as Flow-GRPO are limited in efficiency by poor credit assignment, while forward-process methods such as DiffusionNFT are efficient but unstable without carefully tuned exponential-moving-average (EMA) training of off-policy samples.

The paper says NFT-style objectives contain a hidden forward-reverse consistency force that cancels only partially when positive and negative training are combined, and that EMA stabilizes training through restoration rather than off-policy rollouts. To address this, the authors introduce DiffusionART (Diffusion Reinforcement with Aggregation-First Reward Tilting). It aggregates each prompt's rewards into a shared reward-improving direction for reverse-process states, removes the consistency force by construction, and adds a trust-region-masked second update for drift-adaptive restoration without EMA tuning.

According to the paper, DiffusionART requires no differentiable reward model, remains effective with fewer samples per prompt, and supports pairwise reward models (DiffusionART-Pairwise). Across image and video generation, it trains efficiently and stably without per-task tuning. On HunyuanVideo-13B with equally weighted visual and motion quality rewards, the authors report it reaches the late-stage mean reward levels of DiffusionNFT and Flow-GRPO-Fast with 12.0x and 11.7x fewer GPU hours, respectively, while achieving 28.8% and 53.0% higher mean training rewards. The paper was ranked #1 Paper of the Day on Hugging Face.