MiniMax H3 · Hailuo 03 · powered by Kie
Direct the picture and the soundtrack in one prompt. Create a new MiniMax H3 video from text, animate an opening and ending frame, or guide the shot with up to nine image references.
MiniMax H3 · native stereo
Direct the whole audiovisual shot
Workflow
Text
Output
768P
Credits
180
4–15 seconds · 24 FPS · native stereo audio · powered by Kie
A video stage, not a demo mockup
Choose Text, Frames, or References; set 4–15 seconds and 768P or 2K; then generate a real clip with synchronized stereo audio.
MiniMax H3, also called Hailuo 03, is a general-purpose omni-modal generation system from MiniMax. Instead of treating video and sound as separate finishing steps, it interprets text, images, video, and audio as one creative context and predicts moving pictures with synchronized stereo audio. The official model card lists 4–15 second output, 24 FPS, 32 kHz stereo audio, broad landscape and portrait ratios, and a 2K regeneration workflow. For creators, the practical difference is control: one direction can describe the subject, action, camera path, dialogue, ambience, music, sound effects, timing, and the details that must remain unchanged. This page connects those capabilities to Kie's hosted MiniMax H3 API, so no local model setup is required.
The strongest reason to choose MiniMax-H3 is not one isolated benchmark. It is the ability to combine shot direction, visual references, timing, and sound inside one generation brief.
Direct dialogue, footsteps, room tone, environmental sound, effects, and music beside visible action. MiniMax H3 generates 32 kHz stereo audio with the video, which makes it useful for social ads, product films, dialogue tests, and cinematic previsualization where timing matters. A clear prompt should say what the audience hears, when it enters, and which visible event it must match.
Text mode is the cleanest way to explore an idea. Frames mode adds an exact opening image and an optional ending image for transition control. References mode lets you assign image roles in the prompt, such as using Image 1 for identity, Image 2 for wardrobe, and Image 3 for the location. The modes solve different production problems, so the interface keeps them separate.
Choose any whole-second duration from 4 through 15. Text and reference generation support 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16; reference mode can also adapt to the uploaded material. That range covers cinematic widescreen, standard campaign video, square feeds, vertical shorts, and portrait placements without forcing a crop after generation.
Use 768P while testing composition, motion, and prompt structure, then move to 2K when the creative direction is settled. MiniMax describes its 2K path as in-context regeneration: the lower-resolution result is regenerated with the original creative context instead of being passed through a conventional detail-only upscaler. This is designed to recover small details with awareness of the original scene.
These examples are designed around jobs where coordinated motion, continuity, and sound can reduce the amount of manual assembly after generation.
Build a 10-second hero film from a product reference. Lock the object's proportions and materials, specify a measured camera move, and synchronize a tactile click or musical accent with the final lighting change. This is a practical starting point for landing-page loops, launch teasers, and paid social variants.
Use character, wardrobe, and location images as separate references, then name each role in the prompt. Describe restrained body motion, fabric weight, camera distance, and ambience to test whether a campaign character can remain recognizable through a complete shot rather than a static portrait.
Provide an opening pack shot and a designed end frame, then describe the physical path between them. This workflow is useful for packaging reveals, space transformations, title-card arrivals, and storyboard transitions where both ends of the shot are already approved.
Write the prompt as a short timeline: establish the location, introduce one action, let the camera respond, and resolve on a stable final beat. Add rain, room tone, distant traffic, dialogue, or music as timed story elements. The result can function as a previsualization clip or a finished short-form insert.
A disciplined brief usually produces a more useful first result than a long list of disconnected adjectives. Work from the production decision outward.
Use Text when the composition is open. Use Frames when the first image, final image, or both are already decided. Use References when identity, wardrobe, product design, setting, or style must come from several images. In reference prompts, call the uploads Image 1, Image 2, and so on in their displayed order.
State the subject and setting first, then divide the duration into beats. Add action, camera movement, light changes, and audio cues to the relevant beat. Finish with continuity rules and a short negative list. For a 12-second clip, three or four time blocks are usually easier to follow than twelve separate instructions.
Use 768P to test whether the action, identity, camera, and soundtrack work together. Change one variable at a time. When the take is structurally right, keep the prompt and references stable and switch to 2K. This reduces the cost of exploring directions that will be discarded.
Watch the full clip with sound. Check faces, hands, product geometry, floor contact, reflections, text, cuts, camera continuity, lip sync, and whether audio cues land on the intended frames. A visually sharp clip with an identity drift or mistimed sound is usually less usable than a slightly simpler take that remains coherent.
All three can produce commercial short-form video, but they optimize different creative decisions. Choose MiniMax H3 when a dense multimodal brief and unified stereo soundtrack matter; Seedance 2.0 when natural human motion and action timing lead the brief; or Kling 3.0 when explicit multi-shot structure and a 4K option are the priority. Capabilities shown here reflect current provider documentation and can change.
| Decision | MiniMax H3 | Seedance 2.0 | Kling 3.0 |
|---|---|---|---|
| Best fit | Unified multimodal direction, references, and native stereo audio | Human motion, action, lip-sync, and beat-aware audiovisual work | Multi-shot storytelling, element references, and higher-resolution delivery |
| Hosted input modes | Text, first/last frames, and image/video/audio references | Text, first/last frames, and multimodal references | Text, first/last frames, and named visual elements |
| Duration | 4–15 seconds | 4–15 seconds | 3–15 seconds |
| Top documented output | Up to 2K | Provider-dependent | Up to 4K mode |
| Audio | Native 32 kHz stereo | Native audio with lip-sync and beat matching | Native audio and optional sound controls |
| Commercial choice | Choose for one coherent audiovisual brief across many reference types | Choose when people, performance, and physical movement carry the concept | Choose for planned shot changes, element reuse, and 4K campaign variants |
Yes. MiniMax H3 is also presented as Hailuo 03. MiniMax uses H3 for the general-purpose omni-modal generation system, while Hailuo is the consumer-facing video product name.
You can generate from a text prompt, animate an opening frame with an optional ending frame, or guide a reference-to-video task with up to nine images. Each mode supports 4–15 second output at 768P or 2K.
Yes. The official specification lists native 32 kHz stereo output. Put dialogue, music, ambience, effects, and their timing in the same prompt as the visual direction.
Kie's text and reference endpoints list 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16. Reference mode also supports adaptive framing. Frame-based generation derives its composition from the uploaded image or images.
The official reference specification allows up to nine images, three videos, and three audio clips, with no more than twelve files in total. This page currently exposes up to nine image references; the first five are included in the base generation price, while images six through nine add 11 credits each. Kie's API supports the broader video and audio reference workflow for developers.
Use 768P for prompt and motion iteration. Switch to 2K after the structure is working and you need more finishing detail. The higher-resolution option uses more credits and usually takes longer.
MiniMax publishes downloadable H3 model weights under its own community license, which is not the same as a standard permissive open-source license. The hosted Context-IR and 2K regeneration modules are not included in the initial weight release. Review the current license before local or commercial deployment.
Start at 768P, direct one continuous take, and move to 2K when the motion, identity, camera, and soundtrack are working together.
Open MiniMax H3 generator