🎬 MiniMax-H3 — multi-reference video, with sound

Upload several references and control each one separately. Every image becomes a labelled subject the prompt can name — the woman from <Subject 1>, wearing <Subject 2>, in <Subject 3> — with its own retention strength. Output carries a natively synchronized 32 kHz stereo track.

Up to 9 images, 1 video clip and 1 audio clip. Weights are bfloat16, unquantized.

⏳ Loading MiniMaxAI/MiniMax-H3 (transformer_ref + VAEs, 77.3 GB). Watch the Space logs.

References

Order matters — labels are numbered top to bottom, and reordering is a different request.

Hold to it
fully_preserved = exactly · weak_reference = mood only
Hold to it
fully_preserved = exactly · weak_reference = mood only
Hold to it
fully_preserved = exactly · weak_reference = mood only

Scene

Canvas
2 14

How this runs. MiniMax-H3 is 195.9 GiB in bf16 and a ZeroGPU Space is evicted at 150 GB of storage, so the pipeline is split: the 61.7 GiB transformer_ref partition and both autoencoders live here, and the 62.14 GiB Qwen3-VL conditioner runs in multimodalart/qwen3vl-conditioner, which this Space calls for every request.

Ceiling. Short edge caps at 768 px, 24 fps, 2–14 s. 2K needs H3-Regenerate-2K, which MiniMax has not open-sourced — paid API only.

Cost. ref2va has no turbo LoRA, so a 28-step request is several minutes of an xlarge booking. Draft on the small canvas at a low step count, then spend the steps on the keeper.

Credit. Machinery (h3_split_blocks.py, the split design, the sizing model) is from multimodalart/minimax-h3-reference. The prompt format follows MiniMax's official h3-prompt-writing skill. Weights are governed by the MiniMax-H3 Community License.