MiniMax H3
Nine pictures, three sounds, one script. It shoots the scene cut for cut, word for word.
Faces, places and voices go in as references. Dialogue, ambience and score come out with the picture.
1–15 s · 480p / 768p · landscape or portrait · with sound
Every line, in their voice
Write the dialogue and the model performs it: lips match the words, and each speaker keeps one voice from the first shot to the last. Add a recording of the voice you want and the character speaks in it. Add a music clip and it plays under the scene.
- Lip-synced dialogue in 11 languages, spoken in the language you write it in
- Up to 3 reference audio clips: a voice to match, a track to play under, a sound to reuse
- Three sound layers, three lines of script: dialogue, soundscape, music
- S1
- The conductor · male, sixties, low, unhurried, slight rasp
- S2
- The cellist · female, twenties, soft, precise · voice from <Audio 1>
Cut on the millisecond
One generation can hold several shots. Name the cut time, the framing and the camera move for each, and the edit comes back the way you wrote it, with the sound continuous across the cuts.
- Push In
- Pull Out
- Zoom In
- Zoom Out
- Pan
- Tilt
- Truck
- Pedestal
- Arc Shot
- Tracking Shot
- Static Shot
- Roll
- Shake
- POV
Tell it what to keep
Each reference gets a job in the script and a retention mark that says how much of it must survive: everything, some of it, only its style. Audio works the same way: copy a track outright, or borrow just the voice.
<Subject 1>The conductorfully_preservedFace, moustache, uniform and cap all kept; only his pose and the lighting are new.
<Subject 2>The cellistfully_preservedFace, hair and coat kept; the cello case stays with her in every shot.
<Subject 3>The dining carpartially_preservedLayout, lamps and wood kept; the lighting is restaged for night.
- <Audio 1>Her voicereference
Only the timbre and delivery are followed; the recording itself is not copied.
- <Audio 2>Cello linepartially_copy
The phrase is reused under the second half of the scene, mixed beneath the dialogue.
- Images
- fully_preserved
- partially_preserved
- attribute_transfer
- weak_reference
- Audio
- fully_copy
- partially_copy
- reference
- weak_reference
Two frames, two sizes
Landscape or portrait, composed for the frame rather than cropped into it. Pick 480p for quick drafts and 768p for the final take.
- 480p
- Drafts and previews
- 768p
- Final takes
Spec sheet
MiniMax H3, also known as Hailuo 03, is MiniMax's video generation model. On Plot Party it takes several reference images and audio clips together with a prompt and generates picture and sound in one pass: consistent characters across shots, lip-synced dialogue, scripted cuts, and music or voices taken from the audio you provide.
- Model
- MiniMax H3 (Hailuo 03) by MiniMax
- Clip length
- 1 to 15 seconds per generation; one generation can contain several shots with scripted cut times
- Resolution
- 480p or 768p, landscape or portrait
- Reference images
- Up to 9 per generation: people, animals, products, places, style frames
- Reference audio
- Up to 3 clips per generation: a voice to match, music to play under, sound to reuse
- Audio output
- Stereo sound generated in the same pass as the picture. Dialogue, soundscape and score are written as three separate parts of the prompt
- Dialogue
- Lip-synced speech in 11 languages: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish
- Prompt
- Plain language in any supported language, or a shot-by-shot script with cut times, speaker tags and dialogue
- Best for
- Character interaction and dialogue, product and ad spots, story beats, short-form video, creative composites of several images and sounds
- Weights
- The H3 base model is released as open weights under the MiniMax H3 Community License
- Access
- Plot Party generator and Canvas; REST API
H3 or Seedance 2.5
Two video models, two jobs. Pick H3 when the scene is built from several pictures and sounds and someone speaks. Pick Seedance 2.5 for one long 1080p take directed from a crowd of references.
- Clip length per generation
- MINIMAX H31 to 15 sSEEDANCE 2.54 to 30 s
- Resolution
- MINIMAX H3480p · 768pSEEDANCE 2.5480p · 720p · 1080p
- Reference images
- MINIMAX H3Up to 9SEEDANCE 2.5Up to 30
- Reference audio
- MINIMAX H3Up to 3 clips: voice, music, soundSEEDANCE 2.5Up to 10 clips
- Reference video
- MINIMAX H3Not on Plot PartySEEDANCE 2.5Up to 10 clips
- Audio
- MINIMAX H3Stereo; dialogue, soundscape and score written as three layersSEEDANCE 2.5Dialogue, ambience and score in one pass
- Dialogue languages
- MINIMAX H311, lip-syncedSEEDANCE 2.5Language of the prompt, lip-synced
- Cuts inside one generation
- MINIMAX H3Several shots with cut times to the millisecondSEEDANCE 2.5One continuous take
- Aspect ratios
- MINIMAX H316:9 · 9:16SEEDANCE 2.5Adaptive · 16:9 · 9:16
- Editing modes
- MINIMAX H3Not availableSEEDANCE 2.5Region edit · extend · first and last frame
- Open weights
- MINIMAX H3Base model, community licenseSEEDANCE 2.5No
| FEATURE | MINIMAX H3 | SEEDANCE 2.5 |
|---|---|---|
| Clip length per generation | 1 to 15 s | 4 to 30 s |
| Resolution | 480p · 768p | 480p · 720p · 1080p |
| Reference images | Up to 9 | Up to 30 |
| Reference audio | Up to 3 clips: voice, music, sound | Up to 10 clips |
| Reference video | Not on Plot Party | Up to 10 clips |
| Audio | Stereo; dialogue, soundscape and score written as three layers | Dialogue, ambience and score in one pass |
| Dialogue languages | 11, lip-synced | Language of the prompt, lip-synced |
| Cuts inside one generation | Several shots with cut times to the millisecond | One continuous take |
| Aspect ratios | 16:9 · 9:16 | Adaptive · 16:9 · 9:16 |
| Editing modes | Not available | Region edit · extend · first and last frame |
| Open weights | Base model, community license | No |
Questions
What is MiniMax H3?
MiniMax H3, also called Hailuo 03, is MiniMax's video generation model. On Plot Party it takes up to 9 reference images and 3 reference audio clips together with a prompt and generates a 1 to 15 second video with sound in one pass, including lip-synced dialogue and several shots with scripted cut times. You run it in the Plot Party generator or Canvas, or through the API.
Is Hailuo 03 the same as MiniMax H3?
Yes. MiniMax H3 is the model's official name and Hailuo 03 is the name it carries in the Hailuo app. They are one model.
How many references can I use?
Up to 9 images and 3 audio clips per generation. Images can be people, animals, products, places or style frames. Audio can be a voice you want a character to speak in, a piece of music to play under the scene, or a sound to reuse. Name each reference in the prompt in the order you uploaded it and say what it is for.
What can the reference audio do?
Three things, depending on what you ask for in the prompt. A recording of a voice makes a character speak in that timbre and delivery. A music clip plays under the scene, in full or in part. A sound can be reused as an effect. You can also state how closely to follow each clip: copy it outright, copy part of it, or reference only its character.
Does MiniMax H3 generate dialogue and lip sync?
Yes. Write the lines in your prompt and the character speaks them, lips matched to the words, in the language you wrote. Speech is supported in 11 languages: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish. Give each speaker a stable tag and their voice stays the same across every shot.
How do the three sound layers work?
Dialogue is written inside the shot description next to the action it belongs to. Ambience and physical sounds such as wheels, wind and footsteps go in a separate soundscape line. Music that only the audience hears goes in a third line, described by instruments and tempo or taken from a reference clip, or left as none. Keeping them apart is what stops sound effects from landing on top of a spoken line.
How long can a MiniMax H3 video be, and at what resolution?
Any length from 1 to 15 seconds. Output is 480p or 768p, in landscape or portrait. If you need a single shot longer than 15 seconds or 1080p output, Seedance 2.5 on Plot Party generates up to 30 seconds in one take.
Can I cut between shots inside one generation?
Yes. Describe each shot in order and give every shot after the first a cut time, for example at 00:06.000. Framing, camera moves and dialogue are written per shot, and the soundscape runs continuously across the cuts. Cut only when the new shot brings new information; for a closer view of the same subject, use a camera move instead.
How does MiniMax H3 keep characters consistent?
Upload one clear reference image per character, product and place, and name it in the prompt. One full-body image per character works better than several crops. You can also mark how much of each reference must survive: kept fully, kept in part, or borrowed only loosely for style. The model holds the references across every shot in the generation.
Is MiniMax H3 open source?
The H3 base model is released as open weights under the MiniMax H3 Community License. On Plot Party you use it without hosting anything.
How is MiniMax H3 billed on Plot Party?
In Plot Points per generation. The rate depends on the clip length and resolution you choose, and current rates and plans are on the pricing page.
Can I use well-known characters or brands in a prompt?
No. Prompts that mention well-known intellectual property are blocked before generation. Use your own characters, products and reference images instead.
Hold on the last frame.
Nine pictures and three sounds in, one scene out: cut where you said, spoken as you wrote it, in the voices you gave it.