MiniMax H3 Prompting Cheat Sheet: Direct a Shot, Don't Describe an Image
zesh61 · reddit · 2026-08-14
This post summarizes practical prompting tips for controlling the MiniMax H3 video model. The core philosophy is: "Don't describe an image. Direct a shot."
Key Structure: Subject + Action + Environment + Camera + Timing + Audio
- Action First: Make changes explicit; the model needs to understand dynamics over time.
- Camera Language: Use specific cinematography terms (e.g., slow dolly in, handheld close-up) to avoid ambiguity.
- Think in Beats: Split complex scenes into timestamped shots (e.g., 0-3s: Wide shot... 3-7s: Push in...) rather than stuffing everything into one paragraph.
- Visible Speaker: If there's dialogue, make it painfully obvious who is speaking and when to avoid off-screen voice issues.
- Separate Audio Layers: Treat dialogue, ambient sound, SFX, and music as distinct layers.
- Motion over Appearance: Use verbs to describe motion rather than just static appearance.
A reusable template covering scene, subject, timestamps, audio, and visual style is provided.
Related event: Optimizing MiniMax H3 Prompts Shared(2 posts)→
More from Multimodal
- Flux 3 Klien first impressions: occasional body horror like SD3, otherwise solid — rionix88 · 2026-10-16
- XGRIDS builds world's largest 3DGS town: 23.5 km² and 6.2 billion Gaussian Splats — bilawalsidhu · 2026-10-16
- IMMERSION: AI film compilation links 12 shorts with one continuous camera move — AI_Andrew · 2026-10-16
- A song with vocals, lyrics and 3D head animation created entirely by Opus 5.5 — repligate · 2026-10-16
- Alibaba open-sources CommerceGen-Studio for commercial-grade e-commerce images — alibabagroup · 2026-10-16
- EVA-Cap: Fudan and OPPO align video captioning with atomic audiovisual events via GRPO — Fudan-University · 2026-10-16