Hailuo Reference to Video: Characters, Motion and Voice

Consistency is the hard part of AI video. A character who changes face between shots, or a product whose label drifts, makes a sequence unusable. Hailuo's earlier models offered a subject reference for a single person; MiniMax H3, the model behind Hailuo 3, widens that into reference generation across images, video and audio. This guide explains what each reference type can carry, the limits MiniMax documents, and how to write prompts that keep the model from mixing them up.

What a reference can carry

MiniMax describes reference generation as referencing "character, motion, camera, style, voice, or editing rhythm". Each input type suits different jobs:

Files have size limits too: images up to 30 MB, videos up to 50 MB, audio up to 15 MB each, with sides between 256 and 5,760 pixels.

Preparing references

• Characters: a sharp, front-facing image in even light, face unobstructed. Add a second angle only if the first does not hold.
• Products: the item alone on a plain background, label readable, no reflections of other objects.
• Locations and mood: a frame that shows light, palette and grain without a strong subject that could leak in.
• Motion and camera: a single continuous clip of the movement, without cuts if you want one take.
• Voice: clean speech, no music, one speaker. Use only voices you have permission to use.

Writing the prompt

MiniMax's examples follow a clear pattern: format, then roles, then the story. A brand film example assigns "Image 1: mood, setting, film grain. Image 2: character. Image 3: bag. Image 4: ending logo" and then tells the story in one paragraph. A dialogue example says "Voice timbre follows reference audio 1". Use the same habits:

1. Open with format: duration and ratio, for example "10s, 16:9".
2. Give every reference a role, numbered in upload order: "Image 1: the character. Image 2: the café interior. Video 1: camera movement."
3. Exclude what should not carry over when a reference contains more than you want: "use only the camera move from Video 1, not its people".
4. Tell one story with beats in order, then light and camera style.
5. Describe sound last, including any line of dialogue in quotation marks and the voice reference to use.

10s, 16:9. Image 1: the character, a cyclist in a yellow jacket. Image 2: the city bridge at dawn. Video 1: camera movement only. The cyclist rides across the bridge, slows, and looks out over the river. Soft early light, light mist. Sound: tyres on wet tarmac, gulls, no music.

Keeping a character across a sequence

1. Build a small reference set for each recurring character and product, and reuse exactly the same files for every shot.
2. Describe the character with the same short phrase in every prompt ("the cyclist in the yellow jacket").
3. Keep the style or mood reference constant across the sequence.
4. Generate drafts at 768p, pick the takes where identity holds, and render finals at 2K.
5. Join shots in an editor, using cuts to hide small differences.

Hailuo's earlier subject-reference feature, available with the 2.x generation, follows the same principle with a single face image; the image to video guide covers frame-based work, and the H3 guide covers durations and resolutions.

Rights and consent

References make it easy to bring real people, voices and brands into a generation, which is exactly why they need care. Use photographs of people only with their permission, and never a recognisable person in a context they have not agreed to. Use voice samples only from speakers who have consented to their voice being used this way. For logos and products, use your own or your client's, with approval. MiniMax's examples include brand logos and product references, but your licence and your client's legal team decide what can be published. Keep a record of every reference file used in each delivered clip; if a question comes up later, it is the only way to show what went into the shot.

Fixing common problems

• Two references blend into one face. Give each its role explicitly, or reduce to one character reference per shot.
• The location reference adds a person. Use a location image without people, or say "use only the setting from Image 2".
• The motion copies the reference's subject. State that the video reference is for camera or motion only.
• The voice sounds off. Use a longer, cleaner sample within the 15-second limit, with no background music.
• Too many inputs, muddled result. Remove references until each one has a clear job; quality usually rises with fewer.

Try it

6s, 9:16. Image 1: the product, a green glass bottle with a cream label. Image 2: mood, a sunlit bathroom shelf with soft shadows. A hand places the bottle on the shelf and turns it so the label faces the camera. Static close-up, morning light. Sound: a soft glass clink, distant birdsong.

Run it with a clean product image and a mood image without products. If the label distorts, crop the product reference tighter and add "keep the label text exactly as in Image 1".