MiniMax H3 Reference-to-Video API by MiniMax
MiniMax H3 reference-to-video: generate a video that keeps the subject from a reference image, driven by a text prompt. Supports 2K, 5-15s.
MiniMax H3 Reference-to-Video は MiniMax が開発したモデルです。Atlas Cloud（運営：Atlas Cloud AI LLC）は本モデルへのアクセスを提供するものであり、その所有者ではありません。すべての商標は各所有者に帰属します。
MiniMax H3 Reference-to-Video
MiniMax H3 Reference-to-Video generates a video that keeps the subjects from your reference materials consistent throughout, driven by your text prompt. Instead of animating a fixed frame, it uses the references as identity anchors — ideal for placing a character, product, or style into new scenes and motions. References can be any mix of images, videos, and an audio track (e.g. a music beat) to sync the motion to.
Why Choose This?
-
Subject consistency Preserve one or several people, characters, or objects' identities across the whole clip.
-
Mix reference types Combine reference images, reference videos, and reference audio in a single request.
-
Audio sync (optional) Provide a reference audio track and sync the action to the beat.
-
Prompt-driven scenes Put the reference subjects into any scene or action you describe.
-
High resolution output Generate videos in 2K quality.
-
Flexible aspect ratios adaptive, 21:9, 16:9, 4:3, 1:1, 3:4, or 9:16.
-
Selectable duration Produce 4–15s clips.
Parameters
| Parameter | Required | Description |
|---|---|---|
| prompt | Yes | Text description of the scene and action for the subjects |
| refers | Yes | Array of reference materials, each { "url": "...", "type": "image|video|audio" }. type is optional (inferred from the URL). At least one image OR video is required; audio alone is not allowed. |
| resolution | Yes | Video resolution. Available options: 768P, 2K |
| duration | Yes | Duration of the generated video in seconds. Integer between 4 and 15. Available options: 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 |
| ratio | No | Aspect ratio of the generated video. Defaults to adaptive (the most suitable ratio is chosen automatically based on the input). You may also explicitly specify a concrete ratio. Available options: adaptive, 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 |
| prompt_expansion | No | Enable AI prompt expansion via H3 Context-IR. When true, your prompt is first expanded into a rich, structured description (shot, soundscape, music) before generation, and an additional per-token Context-IR fee applies. Default false: the prompt is used as-is with no extra charge. |
refers example
"refers": [ { "url": "https://.../subject1.png", "type": "image" }, { "url": "https://.../subject2.png" }, { "url": "https://.../clip.mp4", "type": "video" }, { "url": "https://.../beat.mp3", "type": "audio" } ]
How to Use
- Add your references — any mix of images and/or videos (at least one), optionally plus an audio track.
- Write your prompt — describe the scene, action, and camera movement for the subjects.
- Set resolution and duration — balance quality against generation speed.
- Choose an aspect ratio — match your target platform, or use adaptive.
- Run — submit and download your video.
Best Use Cases
- Character Consistency — Keep the same character across multiple shots.
- Product Placement — Feature a specific product in generated scenes.
- Branded Content — Maintain a consistent mascot or spokesperson.
- Creative Series — Build multi-clip stories around a recurring subject.
Pro Tips
- Use clean, well-lit reference images with each subject clearly visible.
- Give one image per subject when combining several subjects in one video.
- Describe the new scene and action, not the reference materials themselves.
- Supply a reference audio track to drive the timing/beat of the motion.
- Combine with different ratios and durations to repurpose the same subjects across platforms.
Notes
promptand at least one reference image or video are required; a request with only audio is rejected.- Reference-to-video and image-to-video are mutually exclusive: this model uses reference materials, not first/last frames — do not mix the two.
- Ensure all reference URLs (image / video / audio) are publicly accessible.
- Generation is asynchronous — submit, then poll for the finished video.
Related Models
- MiniMax H3 Text-to-Video — Generate video directly from a text prompt.
- MiniMax H3 Image-to-Video — Animate a first-frame image (optionally with a last frame).
