MiniMax: H3

One audiovisual result from a multimodal brief

MiniMax H3 is MiniMax’s omni-modal video system. It interprets text, images, video and audio together, then generates video with native stereo sound. The maker presents it for short creative work such as product visuals, advertising, animated posters and film-title concepts, where the relationships among reference assets matter as much as the scene description.

The documented system produces 4–15-second video at 24 frames per second with 32 kHz stereo audio. Its H3-Base generation operates at 768p; the full 2K workflow regenerates that result using the original multimodal context. This is more than a conventional resize: the second stage uses the model and source context again to recover fine detail. H3 Max and H3-Context-IR have separate roles and limits.

Text and frame-guided creation

Text-only input describes the complete audiovisual timeline. A first image anchors the opening, a last image anchors the ending, and two images can define both. For a first-frame shot, begin with the composition and identity visible in the still, then describe the action developing from it. For a first-to-last-frame transition, describe the observable movement connecting the two states rather than repeat two unrelated image descriptions.

A practical product brief can specify a closed package in the opening frame, the hand that opens it, the camera’s gentle movement and the final state to reach. Preserve the object’s shape, label placement and lighting in the supplied frames. Frame-guided creation and reference-guided creation are separate input modes; the maker’s interface does not mix first/last-frame roles with reference-media roles in one request.

Reference-guided generation and editing

The reference mode accepts text with images, video clips and audio clips. Identify what each source contributes: a character’s appearance, a location, camera motion, the temporal structure of a source clip or a voice. A video used only for camera movement is different from a video that should actually be edited or continued. State that relationship explicitly.

The maker’s reference guide separates subject definitions, a task summary, retention analysis, a detailed shot description, overall soundscape and audience-only background music. It distinguishes preservation of visible content from attribute transfer or a broad style reference. For audio, copying a signal is different from referencing timbre, rhythm or delivery. Keep those roles stable throughout the brief.

For a reference-driven scene, supply a character image, a short camera-movement reference and an audio voice reference, then describe where each should affect the target shot. State which clothing, environment or object details must remain recognizable. Compare the finished clip with those requirements, including motion continuity and the relationship between the visible speaker and the voice heard.

Speech, sound and shot timing

H3 jointly models dialogue, physical sounds, ambience and music. The model card documents stable dialogue support for Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish; other languages are described as varying in support. Provide the exact spoken words and language, with a stable speaker identity across shots.

The maker’s prompt-writing guides use English for descriptive sections while preserving dialogue and visible text in their original language. Spoken content uses <d>[Language] ...</d>, with speaker identity and voice description outside the block. Distinguish physical speech from off-screen voiceover, and identify when a line continues across a cut or is cut off by the end. Preserve visible text exactly as well. Fit the amount of speech to the chosen duration and inspect the rendered words and audible line together.

Describe camera movement as an action, including its range and speed when they matter. Number shots in playback order and keep cut times within the video duration. Separate scene sounds such as footsteps or a door closing from background score audible only to the audience. These choices make a short shot plan easier to interpret and easier to review against its result.

Input preparation and numerical limits

The current generation schema requires non-empty text in every mode, with up to 7,000 characters per text item. Images can be JPG, JPEG, PNG, WebP, HEIC or HEIF, up to 30 MB each, with each side from 256 to 5,760 pixels and aspect ratio from 0.4 to 2.5. Frame input permits one opening and one ending image; reference mode permits up to nine reference images.

Reference video can contain up to three MP4 or MOV clips, each 2–15 seconds and together no more than 15 seconds. A clip can be up to 50 MB; its video codec is H.264 or H.265, with AAC or MP3 audio, sides from 256 to 5,760 pixels, ratios from 0.4 to 2.5 and frame rate from 23.976 to 60. Reference audio supports up to three WAV or MP3 clips, up to 15 MB each, each 2–15 seconds and together no more than 15 seconds. The model card caps mixed reference inputs at 12 files. The documented request-body limit is 64 MB, distinct from the per-file limits.

Choose only references that contribute to the brief and make every contribution explicit. A clear subject view is more useful than several conflicting views with no explanation. Preserve reference order, crop away irrelevant material without removing necessary scene context and check that short source clips contain the action or voice characteristic being requested.

Output shape and the complete system

The manufacturer’s H3 interface offers 768P and 2K. Text-only generation requires a concrete ratio: 21:9, 16:9, 4:3, 1:1, 3:4 or 9:16. First/last-frame creation takes its aspect ratio from the input image; reference creation can use an adaptive or explicit ratio. Integer output duration is 4–15 seconds.

The model card distinguishes H3-Context-IR, which interprets and enriches the brief, H3-Base, which generates audiovisual content, and H3-Regenerate-2K, which creates the higher-resolution result. The released base checkpoints separate frame-guided and omni-reference tasks. The context and regeneration components are distinct from those base weights, so the name of an open checkpoint alone does not establish the whole system’s behavior.

The maker’s generation workflow is asynchronous. Its task-query interface retrieves tasks from the preceding seven days. Keep the task identity until the actual video result is returned, then save the clip rather than retain only a time-limited download address. Review visual identity, small text, object interactions, spoken content and sound timing at the intended viewing size. The manufacturer identifies multimodal understanding and fine visual detail as continuing areas for improvement; a detailed brief and source comparison remain central to a useful result.