MiniMax H3 Omni Reference: plan images, video, and audio

MiniMax H3 Omni Reference combines several media sources inside one video brief. The launch manual allows up to nine images, three video clips, and three audio clips, with a maximum of 12 files in a mixed request. The feature is powerful because each asset can contribute a different part of the result. It becomes difficult when several assets compete for the same role.

The solution is reference planning. Before uploading anything, label what every file is meant to control and how the result will be reviewed.

Supplied H3 demonstration: multiple reference roles are resolved into one edited result.

Build a reference role sheet

Start with identity. Use a clear portrait for facial features and a full-body image for silhouette, clothing, and proportions. If a product is central, give it a separate reference with readable geometry and material detail.

Next, define the environment and visual treatment. A location image can set architecture and spatial layout. A style image can guide color, grain, lighting, or graphic language. Keep those roles separate in the prompt. "Image 3 controls the desert road location; image 4 controls the muted 35mm texture" is clearer than "use images 3 and 4 as style."

Video references can carry action, camera movement, editing rhythm, or an existing scene to modify. Do not assume the model knows which one matters. State whether the character should copy the gesture, the camera should copy the move, or the final video should follow the cut pattern.

Audio can guide dialogue, voice tone, music, or timing. The manual requires an audio reference to accompany at least one image or video.

Stay within the media limits

Each reference video can last from 2 to 15 seconds, and all video references together can total no more than 15 seconds. Audio clips follow the same 2-to-15-second range and a 15-second combined limit. A request may include up to three of each.

Images can be up to 30 MB each, videos 50 MB each, and audio 15 MB each. The mixed file count still cannot exceed 12. For API use, the request body is limited to 64 MB, and the manual recommends URL-based media input.

These limits reward selection. Trim a video reference to the exact action or camera move. Cut an audio clip to the required line or tone. A shorter reference is easier to assign and review.

Prevent reference conflict

Check the package for contradictions. Two identity images should agree on hair, face, and age. Product references should agree on color and shape. A camera reference should not demand a fast orbit when the prompt asks for a static product view.

When a difference is intentional, describe it as a transformation. For example: "Use image 1 for the opening wardrobe and image 2 for the jacket after the smoke transition." Without that instruction, the model receives two incompatible targets.

Review by reference role

After generation, score each reference separately. Did the portrait hold through movement? Did the product retain its form? Did the camera follow the reference without changing the action? Did the audio arrive at the correct beat?

This role-based review is more useful than asking whether the result looks good. It shows which source worked and which one needs to change.

Frequently asked questions

How many files can MiniMax H3 Omni Reference use?

The mixed maximum is 12, within the separate limits of nine images, three videos, and three audio clips.

Can I upload audio without an image or video?

No. The manual says audio must accompany image or video input.

Should I always use the maximum number of references?

No. Use the smallest consistent set that defines the shot and gives every file one clear job.

Create a reference role sheet in ArtArch before generating, then attach only the H3 assets that earn a place in the brief.