Kling's DiffusionART cuts video RL fine-tuning cost by 12x while raising rewards up to 53%

KlingTeam · hf · 2026-10-12

KlingTeam introduces DiffusionART, an online RL method for fine-tuning diffusion models that fixes the efficiency/stability tradeoff between reverse-process methods (Flow-GRPO) and forward-process methods (DiffusionNFT).

The authors show NFT-style objectives contain a hidden forward–reverse consistency force that joint positive–negative training only partially cancels, and that EMA stabilizes training through restoration rather than off-policy rollouts.
DiffusionART aggregates each prompt's rewards into a shared reward-improving direction for reverse-process states (removing the consistency force by construction) and adds a trust-region-masked second update for drift-adaptive restoration — no EMA tuning, no differentiable reward model, and support for pairwise rewards (DiffusionART-Pairwise).
On HunyuanVideo-13B with equally weighted visual and motion quality rewards, it matches late-stage mean rewards of DiffusionNFT and Flow-GRPO-Fast with 12.02x and 11.72x fewer GPU hours, while achieving 28.8% and 53.0% higher mean training rewards.

Original post →

Kling's DiffusionART cuts video RL fine-tuning cost by 12x while raising rewards up to 53%

More from Multimodal