MiniMax-H3-Ref2VA — Video Generation AI Model
MiniMax-H3 Ref2VA (omni-reference) generates a 5-15s 24fps video with native stereo audio from a prompt plus a reference image and a reference clip, that clip's own soundtrack included. Video and its soundtrack come out of one denoising loop, so lip movement and sound land in sync. ~144GB bf16 weights — offloaded component-by-component, minutes-scale per clip.
What it does
MiniMax-H3 Ref2VA turns a prompt plus reference media into a five to fifteen second clip at 24fps with its own stereo soundtrack. Picture and sound come out of a single denoising loop, so they are in step rather than aligned afterwards.
References are what this checkpoint is built around; a prompt on its own will not drive it. Attach an image, a clip, or both, and they guide the subject, the style and the sonic character. Six trained canvas sizes, from widescreen to vertical.
When to use it
- Video and its soundtrack generated together, not in two passes
- A reference image anchoring the subject or the style of a shot
- A reference clip lending the scene its sonic character
- Narrative shots of up to fourteen seconds, not micro-clips
- Six trained canvases, from widescreen through to vertical
Details
- CategoryVideo Generation
- Year2026
- LicenseMiniMax-H3-Community-License
Compliance & Provenance
- ProviderMiniMax (open weights) · Specialized
- EU AI Act RiskLimited Risk
- Art. 50 TransparencyRequired — AI-generated outputs are marked
Inputs & Outputs
- TextInputstring
Describe the scene, the motion and the sound you want, and refer to the references by their tags — '<Picture 1>' for the image, '<Video 1>' for the clip, '<Audio 1>' for that clip's own soundtrack.
- ImageInputimageoptional
Reference image (subject, style or scene). Connect this or Video — each port is optional on its own, but ref2va does not run on the prompt alone. It does not bind the generated geometry — the canvas is the one the resolution parameter selects.
- VideoInputvideooptional
Reference clip for motion and camera. Connect this or Image — each port is optional on its own, but ref2va does not run on the prompt alone. It is conditioned on together with its own soundtrack, which has to be mono or stereo. Anything longer than the clip being generated is truncated to it, so match its length to num_frames — about 5.2s at the default 124 frames. 2s is the shortest clip MiniMax documents and 15s the longest.
- VideoOutputvideo
Generated video with its synchronized stereo soundtrack muxed in
Parameters
- num_framesstringdefault 124 frames (~5.2s @ 24fps)options: 124 frames (~5.2s @ 24fps), 158 frames (~6.6s @ 24fps), +8 more
Clip length (5-15s at 24fps)
Total frames to generate, required for ref2va. Values sit on the '17 * n + 5' grid the video VAE can decode; anything else is snapped up to the next one, and the resulting duration has to stay inside the 5-15 seconds the checkpoint generates. The listed values are the ones MiniMax's own second-based API resolves to for a 5 to 14 second request, so 192 frames is the only one that lands on a whole second (8.000s). 124 frames (~5.2s, default) is the shortest and cheapest. To match a reference soundtrack exactly, compute round(samples / sample_rate * 24) and pick the next value up. Cost grows with length.
- resolutionstringdefault 1344x768 (16:9)options: 1344x768 (16:9), 768x1344 (9:16), 1024x768 (4:3), +3 more
Canvas (W×H) — the model's six trained aspect ratios
Output canvas. The six values are exactly what MiniMax-H3's own shape policy resolves its six documented aspect ratios to: a 768 pixel short edge, capped at a 768 * 1344 area budget, both axes then rounded to a multiple of 32. 21:9 is the one ratio the area cap binds on, which is why it is 1536x672 rather than 1792x768. The reference server takes a short edge and a ratio rather than pixels and rejects any short edge but 768, so these six are the whole set the released checkpoint is served at. 1344x768 (default) is the canvas a request generates on when it asks for nothing — references, unlike a keyframe, do not bind the generated geometry. Cost tracks area: the square 768x768 carries about 43% fewer video tokens than the default.
- stepsintegerdefault 50range 20–75
Diffusion steps (20-75)
Number of denoising steps. Default 50 is the diffusers default for this pipeline. The count is sigma grid points with the terminal 0 included, so it drives one model evaluation less. The checkpoint is guidance-distilled — every step is exactly one forward pass, so cost is linear in this value. Drop toward 20 to preview, raise toward 75 if artifacts remain.
- seedintegerdefault -1range -1–2147483647advanced
Random seed (-1 = random)
Random seed for reproducibility. -1 (default) picks a fresh seed every run. A request draws the reference conditioning noise first, then the video noise, then the audio noise from this one seed, so the same seed with the same references, prompt and parameters returns the same video and the same soundtrack.
Limitations
- A person's likeness with a reference voice is a lip-synced result, and the output has to be disclosed as AI-generated
- One image and one clip: the checkpoint takes more, this node exposes a single port per modality
- Around 144GB of weights, offloaded piece by piece, so a clip takes minutes
Tags
Alternatives in Video Generation
- Cosmos3 Nano
NVIDIA world model for text- and image-to-video, with optional synced audio.
- Helios Base
On-device text-to-video, up to 60s @ 24fps. 3-10 min.
- Helios Distilled
Fast distilled text-to-video, fixed 640x384. 1-5 min.
- Wan2.2 TI2V 5B
Wan2.2 dense 5B unified text+image-to-video. Text only → 5s video from prompt; text + image → image as the first frame of the video.