Mixer AI
MiniMax H3
Build a scene from separate references for appearance, movement and sound.
Updated
When to choose it
MiniMax H3 is useful when several inputs must work together: a photo for identity, a video for performance and audio for sound. You can also start from text or connect a first and last frame. Mixer AI offers 4–15-second clips at 768p and 2K.
- Combine photos, video and audio in one task, assigning each a role.
- Generate stereo sound with the picture and use first and last frames.
Modes and settings
- Text mode, one or two frames, or references.
- Up to 9 images, 3 videos and 3 audio files. Video and audio each have a separate 15-second total limit; audio needs a visual reference.
Price in coins
| Quality / size | Input | Coins / second | Extra image, coins / image |
|---|---|---|---|
| 768p | First / last frame | ≈ 12.1 | — |
| 768p | References without video | ≈ 12.1 | ≈ 6.05 |
| 768p | Text | ≈ 12.1 | — |
| 768p | With video reference | ≈ 12.1 | ≈ 6.05 |
| 2K | First / last frame | ≈ 18.53 | — |
| 2K | References without video | ≈ 18.53 | ≈ 5.7 |
| 2K | Text | ≈ 18.53 | — |
| 2K | With video reference | ≈ 18.53 | ≈ 5.7 |
Billing is per second, not per fixed-length clip. The rate is multiplied by the billable duration; the complete request total is rounded up to a whole coin once.
Without video input, billing uses the output duration. With a video reference, it uses input video time plus output time at the video-reference rate. Photos and audio do not add video seconds.
Up to five images are included. Each image after the fifth adds the amount in the extra-image column once per request, not per second. The surcharge is added before rounding the total.
≈ marks a fractional rate rounded to two decimal places for display; billing uses full precision.
In US dollars: from ≈ $0.133 per second. 1 coin ≈ $0.011.
What you can make
- A product scene with a specified character and camera move.
- Give an existing moment a different visual treatment through references.
How to direct the result
- Specify what each file supplies: face, clothes, motion, light or voice.
- For a transformation, list what stays first, then give one main change.
Example: change the lighting, keep the character
10 seconds. Use the photo for character appearance and the video for movement and composition. Keep the walk, clothing and camera direction. Change the setting to an overcast morning: soft diffuse light, wet pavement, no neon signs. Sound: footsteps and quiet street ambience.
What to keep in mind
- The first five images are included in the base calculation; additional images increase the price.
- Input video adds billable duration. Check the total, not only the resolution.
Sources
Model facts verified: 2026-09-20.
Guides and comparisons using this model
FAQ
Can I supply audio alone?
No. Audio references must accompany an image or video.
How are frames different from references?
Frames anchor the opening and ending. References guide attributes and movement without necessarily appearing as exact frames.
Will 2K fix poor movement?
No. Resolve the scene first; resolution sets output size.
Why does a sixth photo change the total?
Five images are included. Each image after the fifth has the surcharge shown in a separate column. It is charged once per request, not multiplied by video duration.
Can I reference more than fifteen seconds of video?
This workflow has a combined 15-second video limit. Select only the relevant excerpt.
Will every untouched detail stay identical?
That is not guaranteed. Compare faces, clothes, objects and background throughout the output.
What changes the price?
The coin rate per second depends on quality, mode and, for some models, sound. Multiply it by the billable duration. Where input video is charged, its duration also counts. The complete request total is rounded once.
Do I need a separate model subscription?
No. In Mixer AI, you top up one coin balance and use it for the tasks you need. You do not need a separate subscription to this model.