MiniMax Hailuo H3 Highspeed Reference to Video

MiniMax H3 Highspeed Reference to Video

What is MiniMax H3 Highspeed Reference to Video?

MiniMax H3 Highspeed Reference to Video is the reference-driven mode of MiniMax's H3 (Hailuo 3.0) Highspeed tier, the speed-and-quality balanced member of the omni-modal H3 video family. You supply reference media - up to nine images of a character, product or place, plus up to three reference videos and three reference audios - with a prompt, and the model generates a fresh 768P clip, 4 to 15 seconds long, that keeps your referenced subject consistent while it performs the action you describe. Native stereo audio is produced in the same pass as the picture. The Highspeed tier trades the standard tier's 2K ceiling for faster turnaround, the practical default when on-model consistency and speed matter more than maximum resolution.

Key Features

• •Reference consistency: keep a character, product or place recognizable inside a brand-new scene the model composes, from up to nine reference images - not just an animated still.
• •Multi-modal references: steer movement and style with up to three reference videos, and voice or sound with up to three reference audios.
• •Native stereo audio generated alongside the video in one pass.
• •Flexible output: 4 to 15 seconds; aspect ratios from 21:9 to 9:16, or adaptive to follow the references; prompts up to 7000 characters.

Best Use Cases

Reference-to-video is strongest when a subject must stay on-model across many clips: a recurring brand character or mascot, a product in lifestyle b-roll, a presenter whose look must match, or a location kept consistent shot to shot. In testing, a single studio portrait of a distinctively styled scientist drove a freshly composed lab scene in which the same character ran an effervescence experiment, proving the subject transfers into an environment the model builds from scratch. It suits social and ad creative needing many on-brand variations fast, explainer clips, storyboards, and product demos.

Prompt Tips and Output Quality

Supply one to three clean, well-lit reference shots on a plain background, then name the subject as "the X in Image 1". Describe one clear action and a motivated camera move, and keep the scene focused rather than crowded. The confirmed 768P output is sharp and coherent with synchronized audio and no identity drift; raise duration for a fuller action and use adaptive ratio to let framing follow your references.

How to use MiniMax H3 Highspeed Reference to Video

MiniMax H3 Highspeed Reference to Video turns reference media plus a prompt into a fresh 768P clip that keeps your chosen subject consistent. Reach for it when a character, product or place must stay on-model across a scene the model builds for you.

Supplying references

Add up to nine reference images of your subject under Reference Images - one to three clean, well-lit shots on a plain background steer best, because a plain backdrop tells the model the image is a subject reference rather than a starting frame. Add up to three reference videos, 2 to 15 seconds each, to borrow a specific motion or style, and up to three reference audios, 2 to 15 seconds each, to follow a voice or a sound. Send at least one reference of any type.

Writing the prompt

Name your subject explicitly as "the X in Image 1", then describe one clear action and a motivated camera move. Keep the scene focused; the Highspeed tier is tuned for fast, legible motion rather than dense, crowded frames. Prompts can run to 7000 characters, but concrete nouns and a single action beat a long description.

Recommended Settings

• •Reference Images: 1-3 consistent shots of the subject. Tested with one studio portrait, which drove a fully composed lab scene with the character intact.
• •Prompt: subject bound as "Image 1" plus one action and one camera move (tested at 768P, 16:9, 6 seconds).
• •Duration: 4-6 seconds for social loops; up to 15 seconds for a longer action.
• •Resolution: 768P, the only option on this tier; use the standard H3 tier for 2K.
• •Aspect Ratio: 16:9 for landscape, 9:16 for vertical, 1:1 for square, or adaptive to follow the references.

In testing, the stored example produced a clean, coherent clip with synchronized native audio and no identity drift.

Examples to try

Each example below ran end to end at the settings shown - 768P, 16:9 - with the clip length and the reference binding carried straight from the card.

• •Reference consistency (character). Prompt: "The scientist in Image 1 - magenta safety goggles, teal nitrile gloves - drops a white effervescent tablet into a tall glass cylinder of clear water on a brushed-steel lab bench; it fizzes violently, a dense column of bubbles and amber foam surges up the glass and spills over the rim, droplets scattering across the steel. The camera tightens on the overflowing rim." The referenced scientist stays on-model in a lab scene the model composes, fizz and all. Tip: a clean studio portrait on a plain background steers best.
• •Product identity. Prompt: "On a white windowsill beside a short stack of loose papers, the retro desk fan from Image 1 spins up: its three chrome blades blur behind the round wire cage and the head oscillates slowly side to side, lifting and fluttering the edges of the papers. The camera pushes in on the spinning cage." The exact fan - chrome cage, teal base - reappears in a fresh room, blades spinning. Tip: use your own product photo and keep to one action.
• •Multi-image fusion. Prompt: "On a concrete rooftop terrace, the woman from Image 1 - short copper-red bob, charcoal wool turtleneck - lifts the oversized round tortoiseshell sunglasses from Image 2 onto her face, settles them, and turns to the camera with a slight smile. The camera arcs slowly around her." Both references hold at once: the same woman in the exact tortoiseshell sunglasses. Tip: give one clean shot per subject, named Image 1 and Image 2.

Reference-guided prompting

Send at least one reference of any type - up to 9 images, 3 videos and 3 audios (2-15s each). Name each in the prompt ("the X in Image 1", "follow the motion in reference video 1") so the model knows which reference drives appearance, motion or voice.

Also worth trying: "the sneaker from Image 1 turns on a studio turntable"; "the mascot from Image 1 walks in, following the motion in reference video 1."

How is reference-to-video different from image-to-video?

Image-to-video animates a supplied frame; reference-to-video keeps a referenced subject consistent inside a new scene built from your prompt.

How many references can I use?

Up to nine images, three videos and three audios, each reference clip 2 to 15 seconds.

Does it generate sound?

Yes - native stereo audio in the same pass as the video.

What does it output?

768P clips from 4 to 15 seconds.

Can I keep a brand character on-model?

Yes - that is the core use. Supply consistent reference images and name the subject in the prompt.