DiffusionART-Flash open-sourced: RL for diffusion models with 25x fewer training timesteps
YangYou1991 · x · 2026-11-01
Yang You's team open-sourced DiffusionART-Flash, an efficient RL post-training method for diffusion models, along with the full MiniMax-H3 recipe (repo at 208 stars, paper arXiv:2610.12572).
Method
- Policy updates train on only 1–2 high-noise timesteps while rollouts follow the full sampling trajectory
- Comparable reward gains and stability vs full training in their experiments
- On a 50-step schedule training 2 steps: 25x fewer sample-timestep pairs
MiniMax-H3 recipe
- 480P, 5s clips, 10 rollout steps
- Video: train on 2 of the first 4 steps; audio: 2 of the first 3
- Six open-source rewards, normalized within prompt groups and routed by modality
Apache-2.0 licensed, supporting t2i, t2v, i2v and omni; the team promises 3 technical answers and 7 engineering lessons to follow.
More from Multimodal
- Filmmaker pushes back on 'AI movies have no soul' with Halloween short made in invideo — Scobleizer · 2026-11-01
- AI-generated 3D animation made in just 3 hours, creator calls it scary — iamfakhrealam · 2026-11-01
- Midjourney --sref combo that yields Korean-style grainy film posters — sergeantsref · 2026-11-01
- Stunning 2D-to-3D character conversion preserves illustration style and details — azed_ai · 2026-11-01
- Beginner tutorial: MiniMax H3 + Pika Flex LoRA generates stereo-audio video in 60s — No-Sleep-4069 · 2026-11-01
- Pika launches H3 Flex: faster-than-real-time video model, up to 80% cheaper than H3Max — techNmak · 2026-11-01