MiniMax: H3 Max

A fast post-trained H3 variant

MiniMax H3 Max is a post-trained variant of the MiniMax H3 base model, optimized around faster generation, prompt adherence and audiovisual quality. The variant was developed through post-training by fal Research. It retains unified multimodal context and natively synchronized picture and sound. Speed is a design priority rather than a fixed completion time for every prompt, duration or application.

Keep Max distinct from the original H3 system and from the separate Max Turbo variant. MiniMax’s current interface lists Max at 480P or 768P, with 768P as the default, 24 frames per second and an integer duration from 5 to 15 seconds. Its documented Max workflow does not offer the original H3 system’s 2K setting or a four-second output.

From a prompt to a shot

Every generation includes a non-empty text brief, up to 7,000 characters per text item. Describe the subject, initial composition, action and result, then make the camera and sound directions clear. For a short object reveal, identify what the viewer should see first, how the object moves into view and how the shot should end.

A useful five-second brief might describe a ceramic cup on a wooden table, a hand placing a saucer beside it, a small camera push and the sound of ceramic meeting wood. Keep the action small enough to fit the selected duration and inspect whether the returned motion and sound follow the same event.

Frame-guided creation

The manufacturer documents an opening frame, an ending frame or both. A first image establishes where the clip begins; a last image establishes the final visual state. Explain the motion connecting them, including changes in pose, object position, camera or lighting. For a two-frame transition, avoid a brief that introduces an incompatible scene between its endpoints.

Frame-guided output derives its aspect ratio from the input image. For text-only work, choose a concrete ratio from 21:9, 16:9, 4:3, 1:1, 3:4 or 9:16. Reference-guided work can use an adaptive ratio or one of the documented explicit choices. Compose the reference or written scene for the target frame rather than assume the same arrangement works equally well in portrait and landscape.

Multimodal references

MiniMax’s current Max generation schema also documents reference images, video and audio. State what each should influence: character identity, an environment, a motion pattern, a camera trajectory, voice timbre or a sound characteristic. First/last-frame roles and reference-media roles are mutually exclusive in that interface.

For a character shot, identify the appearance reference and the motion reference separately, then provide the target action and spoken line. A voice reference that establishes delivery is different from copying its complete audio signal. Keep the brief explicit about the features to preserve and the changes to make, and review the actual audiovisual clip against each reference role.

Prompt expansion

The Max-specific expansion setting offers disabled, balanced and quality choices, with balanced as the default. Choose disabled when the supplied brief should pass through without expansion; use a documented expansion mode when additional interpretation is desired. Keep the original brief so a change introduced during expansion can be distinguished from the requested content.

Describe sound as part of the scene. Identify physical sounds, ambience and any spoken words, and separate them from audience-only background music. Provide the exact line and the intended speaker rather than asking for unspecified dialogue. Listen to the result as well as watch it; a convincing image does not establish that the spoken wording or timing is correct.

Input preparation and limits

The current manufacturer schema accepts JPG, JPEG, PNG, WebP, HEIC or HEIF images, up to 30 MB each, with sides from 256 to 5,760 pixels and ratios from 0.4 to 2.5. Opening and ending roles each allow one image; reference mode allows up to nine images. Preserve a clear view of the subject and the details that must remain identifiable.

Reference video permits up to three MP4 or MOV clips, each 2–15 seconds and together at most 15 seconds, with a 50 MB per-file limit. It documents H.264 or H.265 video, 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 permits up to three WAV or MP3 clips, each up to 15 MB and 2–15 seconds, together at most 15 seconds. The request body has a separate 64 MB limit.

Use only references that contribute to the shot and identify each contribution in the text. A short, relevant motion clip is easier to use as a direction than a long collection of unrelated events. Keep the intended duration and output resolution explicit so the result can be compared with the same brief.

The finished result

The maker’s asynchronous workflow returns a task identity before the completed video is available. Its task-query interface retrieves tasks from the preceding seven days. Save the actual returned clip and its specification for editing or further review; a task identifier is not the audiovisual artifact. Download addresses are time-limited, so keep the file itself when retaining a selected result.

Review subject continuity, motion, visible text, dialogue and sound timing at the size where the clip will be used. Use concrete feedback for another creative brief: the gesture that changed, the line that was omitted or the framing that needs adjustment. The useful result is a short audiovisual shot tied to the supplied content and its explicit reference relationships.