MiniMax H3 Ref2VA Prompt Guide — Reference Roles & Timed Ads
MiniMax H3 Ref2VA prompt guide for commercial ads: assign jobs to product and talent photos, map to official export fields, then build in the Prompt Builder.
Ref2VA is what most brand and e-commerce teams need: you upload real assets (product, model, logo, mood board) and write a script that says what to keep from each file.
Official capacity: up to 9 images, 3 videos, 3 audio clips. Start with 2–4 images until results are stable — more files = more roles to declare, not automatic quality.
Step 1 — Sort your files before you write
| File type | Typical job in an ad |
|---|---|
| Product photo | Lock shape, label, colorway |
| Model photo | Lock face, skin tone, wardrobe |
| Environment / mood image | Lighting, palette, film grain |
| Logo PNG | End card |
| Short video clip | Camera move only (orbit speed, not subject) |
| Audio clip | Voice timbre reference for dialogue |
Write these jobs in a [References] block before any storytelling.
Step 2 — The commercial Ref2VA template
[References] ← who does what (@image1, @video1…)
[Core idea] ← one paragraph: format, tone, what NOT to do
[Process] ← 0–3s, 3–7s… with camera verbs
Dialogue: ← short or "none"
overall_soundscape: …
non_diegetic_music: …
You do not need to memorize the six official field names — the simplified pattern below maps directly into our Prompt Builder. It outputs the exact subject_definitions, summary, and retention_analysis blocks MiniMax expects.
Guide language → official export fields
Use this playbook in plain English; the Builder compiles to official section names:
| What you write (this guide) | Official export field | In Prompt Builder |
|---|---|---|
[References] — job per file | subject_definitions | References (Subject / Picture / Video / Audio) |
[Core idea] — format & tone | summary | Summary (+ task type prefix) |
| How much to keep from each ref | retention_analysis | Retention (fully_preserved, weak_reference, …) |
[Process] — timed shots | detailed_description | Shots ([Shot 1] … At MM:SS.mmm) |
| Ambience, foley, dialogue | overall_soundscape | Sound → overall_soundscape |
| Score / music bed | non_diegetic_music | Sound → non_diegetic_music |
Label tip: guides often say @image1; the official export uses <Picture 1>, <Subject 1>, etc. The Builder generates those labels for you.
Step 3 — Language that locks the product
Vague: “Use the bag from image 3.”
Specific: “@image3 locks the black structured bag — handle curve, gold clasp, logo placement. Do not redesign the bag in later shots.”
Same for faces: “@image2 locks jawline and hairstyle — do not swap identity in close-ups.”
Step 4 — Motion from video references
Community pattern that works:
@video1 provides slow 360° orbit speed and radius only — ignore all subjects and objects in the source clip.
Otherwise H3 may copy unwanted people or products from your reference video.
Step 5 — Pace like an editor
For 10–15s ads, name your cuts:
- “Hard cut on downbeat at 3s”
- “Beat-synced montage 6–8s, 0.5s per insert”
- “Hold logo 2s, fade to black”
Fashion and sport spots fail when the prompt reads like one paragraph with no timing.
Quick checklist before you hit generate
- Every uploaded file mentioned with a job, not just “attached”
- Product named again inside the timeline (“same bottle as @image1”)
- Negative tone guard (“not horror, not comedy trailer”)
- Sound and music in separate lines
- Dialogue fits the duration (one line, maybe two)
Templates to start from
- Desert fashion campaign — multi-image brand story
- Luxury perfume — product lock across locations
- Eyewear runway — vertical social ad
Each template page explains structure → adapt → pitfalls before the copy button.
Keyframe modes (I2VA / FL2VA)
When you already have start or end frames instead of a full reference set:
- I2VA generation — one opening image; describe what changes forward from that frame.
- FL2VA generation — first and last frame; describe continuous motion between them (not a montage).
Upload your frame URLs on the Generate page, paste a short motion brief, and generate a video — or compile a full Ref2VA export in Studio when you need retention analysis and six-section control.
Build your Ref2VA export
Ready to compile? Open the Prompt Builder on the homepage — Ref2VA mode is pre-selected. Upload references (sign in), fill Summary / Retention / Shots, and copy the official export to Hailuo, the MiniMax API, or your workflow.
Open Prompt Builder (Ref2VA) → · Official Hub skills checklist · Browse commercial templates · Back to all guides