Hailuo 3 (MiniMax H3) Guide: Modes, Specs and Prompts

Hailuo, MiniMax's video app at hailuoai.video, built its reputation on natural human motion with the 01 and 02 models and the 2.3 updates. MiniMax H3, which creators call Hailuo 3, is a broader step: one model that takes text, images, video and audio as input, generates sound with the picture and edits existing footage by instruction. This guide explains what H3 does, using MiniMax's own documentation for the model, and how to prompt it. The app may present the same abilities with friendlier labels and plan-based limits; where they differ, follow what your screen shows.

The three ways to generate

MiniMax groups H3's abilities under three headings: native multimodal understanding and generation, precise multimodal editing and control, and production-ready content across film, advertising, e-commerce and games. In practice that means one model can animate a still, keep a character from a reference, follow the timing of a reference clip, use a reference voice and make targeted edits.

Specs that decide your settings

• Duration: 4 to 15 seconds in whole seconds (H3 Max: 5 to 15).
• Resolution: 768p or 2K (H3 Max: 480p or 768p). MiniMax also offers a regeneration step that turns a 768p H3 result into 2K.
• Aspect ratio: common ratios, or adaptive. With a first frame, the image sets the ratio.
• Frames: zero, one or two images; sides between 256 and 5,760 pixels; ratios from 2:5 to 5:2.
• References: up to 9 images, up to 3 video clips (2 to 15 seconds each, 15 in total) and up to 3 audio clips (15 seconds in total), at most 12 files combined.
• Prompt: up to 7,000 characters.

Those limits are long enough for a full 15-second scene with several beats, which changes how you write prompts compared with older 6-second models.

Prompting H3

MiniMax's own examples show a pattern worth copying: state the format, assign each reference a role, then tell the story in order.

15s, 16:9. Image 1: mood and setting. Image 2: the character. Story: she walks into the empty market at dawn, stops at a stall, picks up an orange and smiles at the vendor. Handheld, close focus, warm low sun. Sound: distant traffic, crates being stacked, no music.

1. Format first. Duration and aspect ratio at the start remove guesswork.
2. Name each reference's job. "Image 2: the character" stops the model blending a location into a face.
3. One story, in order. Beats separated by commas or full stops read as a sequence within the clip.
4. Camera and light. One camera style for the whole clip unless you ask for cuts.
5. Sound last. H3 generates audio, so describe ambience, effects and any spoken line; say "no music" if you will score it yourself.

For editing an existing clip, write the change as an instruction, as MiniMax's image-to-video example does: "Pull focus to the people in the background and add more steam to the ramen bowl."

Where H3 sits among Hailuo's models

The 2.x models (02 and 2.3, with Fast variants) are still offered on several platforms and remain good at everyday image to video; Leonardo, for instance, suggested Hailuo 2.3 as the replacement for its own retired Motion 1.0. H3 adds the longer durations, 2K output, audio, references and editing. A practical split: 2.x or H3 Max for quick, cheap drafts; H3 for finished shots, dialogue and anything that needs references. Our Hailuo prompting guide covers the motion vocabulary that applies to every version.

Editing an existing clip

H3's editing is one of its biggest changes. You supply a video and describe the change in plain words, and the model alters what you named while keeping the rest of the footage: the focus, an object, the weather, a sound, the pacing. MiniMax groups this under "precise multimodal editing and control" and says it is meant to make iteration predictable. Use it for small, specific changes rather than a complete rewrite of the shot, and describe what must stay the same as well as what should change. A short test on a 5-second section is cheaper than editing a full 15-second clip, and it tells you quickly whether the instruction is being read the way you meant it.

Common mistakes

• Treating 15 seconds as one action. Long clips need beats; describe them in order.
• Unassigned references. Without roles, images blend.
• Rendering 2K too early. Draft at 768p and upgrade only the keeper.

Settings at a glance

MiniMax also publishes H3 as an open-weight model that can run locally, for example in ComfyUI; that matters mainly to studios with their own GPUs. For references in depth, see Hailuo reference to video.

Try it

8s, 9:16. Image 1: the character, a young chef in a white jacket. She tastes the sauce from a wooden spoon, pauses, then nods and smiles at the camera. Static close-up, warm kitchen light. Sound: a simmering pan, a soft clink of the spoon, no music.

Generate it at 768p first with a single character image as the reference. When the performance is right, rerun or regenerate the result at 2K. Then try the same prompt without the reference to see how much identity the image is holding.