MiniMax H3 Reference to Video

Turn multimodal assets into one coherent MiniMax H3 direction by giving every asset an explicit, bounded job. Stay prompt-only and never submit a generation job.

Confirm this workflow

Use this skill when media acts as a reference rather than an exact first or last frame. If the only visual input is the literal opening frame, use ../minimax-h3-frame-to-video/SKILL.md. If an existing video is being changed locally while the rest stays stable, use ../minimax-h3-video-editor/SKILL.md.

Recommend a 5–15 second clip, 768P for iteration or 2K for a final pass when the active workflow and hardware support it, and a fixed or adaptive aspect ratio. As a conservative complexity budget, prefer no more than 9 images, 3 videos, 3 audio clips, or 12 files total; keep reference video and audio clips within 2–15 seconds and avoid audio as the only reference type. Let the executor validate the installed workflow rather than treating these working limits as an API contract.

Create an asset ledger first

For every supplied asset, record:

• exact label: Image 1, Video 1, Audio 1, and so on;
• positive job: identity, wardrobe, product geometry, environment, style, storyboard, motion, camera, edit rhythm, voice, music, or sound design;
• preservation details that define success;
• excluded influence: what must not be copied from that asset.