MiniMax Hailuo H3 Lite Reference to Video

MiniMax H3 Lite Reference to Video (Image-to-Video)

MiniMax H3 Lite Reference to Video turns your own reference images, clips and audio into a new, moving shot that keeps the subject you supplied. It is the lightweight Lite tier of MiniMax's H3 family, built for exploration and high-volume batch generation.

What is MiniMax H3 Lite Reference to Video?

It is a reference-conditioned video generator. Feed it up to nine reference images of a character, product or place — plus optional reference video and audio — and describe the scene and motion you want. The model generates a fresh 480P or 768P clip, 4 to 15 seconds long, in which the referenced subject keeps its identity, wardrobe and look while performing the action in your prompt. Drafts made here translate predictably when re-run on higher-fidelity H3 tiers.

Key Features

• •Reference consistency — carry a face, outfit, product or location from stills into a generated shot.
• •Flexible references — up to 9 images, 3 reference videos (motion or style) and 3 reference audios (a voice to follow).
• •Strong instruction following — the named subject, setting, camera move and motion appear as written.
• •Format control — 480P or 768P, 4–15 seconds, and aspect ratios 16:9, 9:16, 1:1, 4:3, 3:4 and 21:9.

Best Use Cases

In testing, a single full-body reference of a model in a crimson silk gown was carried into a clean catwalk clip: her face and dress stayed consistent as she strode and the silk train billowed. That suits fashion and product promos, character-consistent social content, storyboard and A/B shot drafts, branded e-commerce clips, and batch variations of one subject before a final high-resolution pass.

Prompt Tips and Output Quality

Reference the input in your prompt (for example, "the woman from Image 1"), then name one clear, verb-driven action. Use one clean, evenly lit reference with the full subject visible for the strongest identity lock. 768P gives the cleanest result; 480P is for faster drafts. Keep the subject large in frame — dense crowds and tiny on-screen text are harder on the Lite tier.

Using MiniMax H3 Lite Reference to Video

MiniMax H3 Lite Reference to Video generates a new clip that keeps a subject you supply as a reference. Start with your references, then describe the motion you want.

Prepare your references

Use one clean reference image with the full subject visible — a clear face for a character, the whole product for a prop, a wide view for a place. The first references carry the most identity weight; up to nine images are supported, plus up to three reference videos (for motion or style) and three reference audio clips (a voice to follow). At least one reference is required.

Write the prompt

Name the referenced subject explicitly ("the woman from Image 1"), then give one clear, verb-driven action and the setting. Keep it concrete: who, where, doing what. Prompts can run up to 7000 characters, but a tight, specific description reads most reliably.

Choose output settings

Pick an aspect ratio that matches the destination — 16:9 for landscape, 9:16 for vertical social, 1:1 for square. Set duration between 4 and 15 seconds; 6 seconds fits a single action cleanly, while longer durations leave room for a full beat. Use 768P for the cleanest result and 480P for faster drafts.

Recommended Settings

The configuration that produced a clean, consistent result in testing:

• •reference_images: one full-body studio shot with a clear face and the full outfit
• •prompt: reference the image, then one action (a stride, a turn, a pivot)
• •ratio: 16:9
• •duration: 6
• •resolution: 768P

768P is the Lite tier's top quality; keep the subject large in frame and lit evenly for the strongest identity lock. Dense crowds and tiny on-screen text are harder on this tier, so lean on a single, well-defined subject.

Examples to try

Each example ran end to end at the settings shown; the subject's identity comes from the reference image and the prompt adds the motion and setting.

• •Character identity (768P, 6s). Prompt: "The woman from Reference Image 1, in her floor-length crimson silk gown with the long train, strides down a glossy black catwalk. The silk billows and snaps with each step, and she pivots hard at the end so the train whips across the polished floor. Overhead spotlights rake the runway and seated spectators line both sides." The reference model's exact face and crimson gown carry into a moving catwalk shot. Tip: use one clean full-body studio shot with the face and whole outfit visible.
• •Place consistency (480P, 4s). Prompt: "The grand opera house interior from Reference Image 1 — its white marble staircase, gilded balconies and rows of red velvet seats — holds steady while the camera glides forward through the empty foyer and tilts up to the glowing crystal chandelier." The referenced hall keeps its architecture as the camera travels through it. Tip: give a wide, evenly lit shot of the whole space and name the features that must stay.
• •Product consistency (480P, 4s). Prompt: "The three-tier dark chocolate cake from Reference Image 1 — its glossy ganache, ring of chocolate curls and gold-leaf flecks — stays exactly as shown while molten chocolate is poured over the top tier and streams down the sides onto the white marble stand." The cake's exact design holds while the pour plays over it. Tip: shoot the product plainly on a clean background and name the action.

Reference-guided inputs

These examples use one reference image. The model takes up to nine reference images, plus short reference videos (motion or style) and reference audio (a voice to follow); address each by number, as in "Reference Image 1".

Further prompts it is built for: "the sneaker from Reference Image 1 spins on a matte turntable", or "the mascot from Reference Image 1 jogs toward the camera."

What inputs does it take?

Reference images (up to 9), reference videos (up to 3) and reference audio (up to 3), plus a text prompt. At least one reference is required.

What resolutions and durations are supported?

480P or 768P (no 2K on Lite), from 4 to 15 seconds.

Will it keep my character consistent?

Yes — supply a clear reference image of the character, product or place you want carried into the shot.

How is it different from the H3 Lite text-to-video and image-to-video models?

Text-to-video builds the scene from words alone; image-to-video animates a supplied first or last frame; this model follows references to keep a subject, motion or voice consistent.