How MiniMax H3 Prompts Actually Work (Plain English)

Most people arrive with a text-to-image habit: one pretty sentence, hit generate, hope for the best. MiniMax H3 is different. It outputs up to 15 seconds of 2K video with built-in stereo sound — dialogue, foley, ambience, music — from a single request. That means your prompt is closer to a short production script than a Stable Diffusion tag list.

We read MiniMax’s official manual, their GitHub examples, and early community tests. Below is the compressed version in normal language — official terms appear only where they help you map back to the API.

---

The one-sentence mental model

Describe what happens on screen, second by second, and what the microphone should hear — in that order.

If you only describe “cinematic luxury vibe,” H3 still generates something — but camera, pacing, and audio will be guesses.

---

Pick your mode first (this saves hours)

| You have… | Use this mode | What you write | |-----------|---------------|----------------| | Nothing but words | T2VA (text-to-video) | Shot list + sound + music | | One hero photo | I2VA | “Start from this frame, then…” + shot list | | First & last frame | FL2VA | Path between those frames — usually no cuts | | Product shots, talent, motion refs | Ref2VA | “Image 1 = face, Image 2 = product…” then timeline |

Rule of thumb for ads: if you already have pack shots or model photos, Ref2VA almost always beats pure text.

---

The minimum that works (text-only)

For T2VA / I2VA / FL2VA, you need three blocks. Official docs label them integrated_multimodal_description, overall_soundscape, and non_diegetic_music — we’ll call them:

1. Shot list — what the camera sees and when ([Shot 1] 0–3s: …)
2. Sound in the scene — rain, footsteps, dialogue, engine — things characters would hear
3. Background music — score the viewer hears; write N/A if you want natural sound only

Floor length: subject + place + action + style. Prompts under ~200 characters are a known failure mode in the official guide — not “minimalist art.”

---

The minimum that works (with reference files)

Ref2VA adds one non-negotiable step before the shot list:

Tell H3 what job each uploaded file has.

• “@image1 = lock the bottle label”
• “@video1 = copy orbit speed only, ignore the person in the video”

Unlabeled uploads are the #1 reason “it ignored my product photo.”

Then write: References → Core idea → Timed process → Sound → Music.

---

Weak vs strong (same idea)

Weak

Beautiful cinematic perfume ad, luxury feel, 4K.

Strong

0–3s: bottle rises on black marble, side light. 3–6s: hard cut — same bottle on terrace at sunset, label identical. Sound: glass tap, evening crickets. Music: slow piano, one chord at 6s.

The second version gives H3 decisions, not adjectives.

---

What to read next

1. Ref2VA with product & talent photos — commercial playbook
2. T2VA shot lists — when you have no assets
3. Sound that doesn’t suck — where to put dialogue vs music
4. Common mistakes — community + official failure patterns

Then open a commercial template, read How this prompt is built, and edit — don’t blind copy.