Running Online RL on MiniMax-H3 Joint Audio-Video Generation with DiffusionNFT
青稞AI · wechat · 2026-09-08
The VeRL-Omni team documents running a complete online RL loop for MiniMax-H3 — an open-weights joint audio-video generation model (up to 15s, 2K, 24FPS, stereo audio; Elo 1227, top-3 on Text-to-Video Arena and the only open-weights model there).
- Using DiffusionNFT, they close the loop for T2VA (text-to-audio-video) and FL2VA (first/last-frame-conditioned) training: vLLM-Omni handles rollout, Diffusers/FSDP2 trains the actor, and VeRL-Omni wires data, rewards, NFT loss, and LoRA old-policy syncing together.
- Rewards combine CLAP (text-audio alignment) and ImageBind (audio-video consistency) to close optimization gaps; key pitfalls include H3's CFG distillation, data-fraction timesteps, and inverted velocity sign, where silent mismatches can hide behind healthy-looking loss curves.
- On 8 GPUs, core reward mean rose from 0.41 to 0.51 in 3+ steps with stable gradnorm (0.06–0.12); a full reproducible config is provided.
More from Multimodal
- ostris pre-training "Primal VAE": 1-channel latents match modern 8x VAEs via patching — ostrisai · 2026-10-26
- Slopcore: AI-generated meme films turn low-fi slop aesthetics into a subculture — Yamapama · 2026-10-26
- TischEins, a large Midjourney SREF library, relaunches with new design after security outage — tisch_eins · 2026-10-26
- MOSS-Image-SFT 4-step distillation tops 16.5K downloads on Hugging Face — _akhaliq · 2026-10-26
- Claude Code can now edit videos: a 10-prompt workflow from raw footage to final cut — TawohAwa · 2026-10-26
- One-sentence 3D worlds turn cinematic with CineMap and TapNow — gizakdag · 2026-10-26