MiniMax H3 Max R2V (Reference to Video)

Overview

H3 Max R2V is a reference-to-video generator by MiniMax, delivered as H3 Max. It takes your prompt plus optional reference images, clips, and audio. The model uses those references as conditioning to carry identity, style, motion, and voice into a new scene. It outputs a short MP4 video with an audio track, so you can iterate on shots fast.

What you can build

• Character-consistent short scenes from a small “character sheet” of images
• Product shots that match a brand style guide and material textures
• Motion-guided animations that borrow camera rhythm from a reference clip
• Video variants that keep the same subject, outfit, or prop across multiple takes
• Prompted scenes with ambience or voice cues from an audio reference

Inputs

• A required written description of the shot you want. Name your references by order, like “Image 1” or “Video 2”.
• Optional reference images for subject identity or style. You can attach up to 9 images. Use them as “Image 1”, “Image 2”, and so on.
• Optional reference video clips for motion guidance. You can attach up to 3 clips. Each clip must be 2–15 seconds. Combined reference video time can’t exceed 15 seconds.
• Optional reference audio clips for voice or sound cues. You can attach up to 3 clips. Each clip must be 2–15 seconds. Combined reference audio time can’t exceed 15 seconds.
• A video duration you choose in whole seconds, from 5 to 15 seconds.
• A target output resolution choice of 480P or 768P.
• An optional aspect ratio choice. You can set a fixed ratio or let the model adapt based on your inputs.
• An optional prompt expansion choice that controls how much the system rewrites your text before generation.

Outputs

• One generated MP4 video file.
• The MP4 includes an audio track generated with the video.
• Some deployments also return helpful metadata such as the final expanded prompt and a seed value for reruns.

Recommended settings

• Use 768P when identity details matter, like faces, logos, or small props.
• Use 480P for rough drafts and fast prompt iteration.
• Pick an explicit aspect ratio when you already know the delivery format.
• Keep prompt expansion on a balanced mode for general work. Disable it when wording must stay exact.

Limitations

• Output length is limited to 5–15 seconds per generation.
• You can attach up to 12 total reference files across images, videos, and audio.
• Audio references can’t be used by themselves. Pair them with at least 1 image or 1 video.
• Reference-heavy requests can conflict. Too many constraints can reduce prompt fidelity.
• Low-quality inputs can degrade results. This includes blurry images, heavy compression, and noisy audio.
• If your references have mixed lighting, mismatched wardrobe, or inconsistent faces, identity may drift.
• The output audio is baked into the MP4. You typically can’t separate clean stems.

Safety & compliance

• Only upload content you have rights to use, including faces, voices, and copyrighted material.
• Don’t use the model to impersonate real people without consent.
• Avoid content that exploits minors or depicts sexual content involving minors.
• Follow applicable laws for synthetic media. Add clear AI-generated disclosure when required.
• If you plan to self-host MiniMax H3 open weights, review the MiniMax H3 Community License and its territory limits and Acceptable Use Policy.