DiffusionART: NUS, Kling AI and HKUST stabilize reward-guided diffusion RL training
jiqizhixin · x · 2026-11-14
Researchers from the National University of Singapore, Kling AI and HKUST released DiffusionART, an arXiv preprint tackling unstable RL training for image and video diffusion models.
Bottleneck: regression-based objectives mix useful reward guidance with an unwanted consistency force; stabilizing via EMA often needs task-specific tuning.
Mechanism: aggregate rewards across generations for the same prompt before guiding actual denoising states; two updates on separate minibatches add a restoring force as the model drifts, with a trust-region mask limiting excessive movement. No gradients through the reward model required.
Payoff: on HunyuanVideo-13B with equally weighted visual and motion quality rewards, training is stabilized with improved quality.
More from Multimodal
- 60-second Rube Goldberg demo built with custom CUDA path tracer and MuJoCo physics — AI_Andrew · 2026-11-14
- mitsuhiko shows Fable turning one landscape photo into stunning hand-painted art — mitsuhiko · 2026-11-14
- Redditor shares "Returning Home", a short AI film made with Seedance — purplenurpleSEB · 2026-11-14
- Fable 5.2 directs a full night scene from one idea, shot in one take by Seedance 2.5 — socialwithaayan · 2026-11-14
- HeyGen's video model impresses in first test: faster than Grok, strong anime results — azed_ai · 2026-11-14
- End-to-end Japanese multi-speaker conversation model trained on 130k+ hours runs on consumer GPUs — kastnerkyle · 2026-11-14