Reinforcement-learning fine-tuning for video diffusion models has always been a painful GPU-hours bill: rewards go up, and the compute budget burns down with them. DiffusionART, a paper submitted to arXiv on October 8, offers a new recipe: the authors report that on HunyuanVideo-13B it reaches the late-stage mean reward level of DiffusionNFT with 12.02x fewer GPU hours, while scoring 28.8% higher mean training rewards. On October 12 the paper topped the Hugging Face Daily Papers listing with 151 upvotes.
Background: Two Routes, Two Traps
Online RL fine-tuning of diffusion models with black-box rewards has split into two camps. Reverse-process methods, exemplified by Flow-GRPO, are limited in efficiency by poor credit assignment, as the paper points out. Forward-process methods, exemplified by DiffusionNFT, are efficient but unstable without carefully tuned exponential-moving-average (EMA) off-policy training.
The paper's central insight: NFT-style objectives hide a forward-reverse consistency force that joint positive-negative training only partially cancels, and EMA stabilizes training through restoration rather than off-policy rollouts.
The DiffusionART Recipe
DiffusionART stands for Diffusion Reinforcement with Aggregation-First Reward Tilting. It first aggregates all sampled rewards for each prompt into a shared reward-improving direction, then applies that direction to reverse-process states — the consistency force is removed by construction instead of being fought during training. On top of this, it adds a trust-region-masked second update for drift-adaptive restoration, so the whole pipeline no longer needs EMA tuning.
Three engineering details stand out. First, no differentiable reward model is required — black-box rewards suffice. Second, it stays effective with fewer samples per prompt. Third, it supports pairwise reward models via the DiffusionART-Pairwise variant. Across image and video generation tasks, the authors report efficient and stable training without per-task tuning. Code is available in the GitHub repository.
The Scorecard: A 12x Gap in GPU Hours
On HunyuanVideo-13B, with equally weighted visual and motion quality rewards: versus DiffusionNFT, DiffusionART reaches its late-stage mean reward level with 12.02x fewer GPU hours and 28.8% higher mean training rewards; versus Flow-GRPO-Fast, it needs 11.72x fewer GPU hours and scores 53.0% higher.
For video-generation teams the implication is blunt: the same compute budget now buys roughly an order of magnitude more iterations. When reward models are black boxes and compute is the hard constraint, differences in training-method efficiency compound into differences in product iteration speed.
The Hugging Face Daily Papers page lists the paper under KlingAI Research. What the community is watching is equally clear: no EMA tuning and no differentiable reward means the engineering barrier to adoption drops sharply.
So the thing to watch is not any single "12x" — it is the curve: the unit cost of RL fine-tuning for diffusion models is falling fast. When fine-tuning becomes cheap enough to do casually, differentiation shifts from "who owns a big cluster" to "who owns a good reward".