MiniMax H3 Reference to Video
Direct MiniMax H3 (Hailuo 03) with the material you already have: up to 9 reference images, 3 reference video clips, and 3 audio tracks in a single generation. Keep subjects and products consistent across shots, borrow motion and pacing from reference footage, sync action to a music track, and get native audio in the output — 5-15 seconds, up to 4K.
50% less than the priciest provider (4 compared). Same model, same output.
Up to 9 reference images for subjects, products, style or composition. Cite in the prompt as Image 1, Image 2.
Up to 3 reference clips (2-15s each, combined up to 15s) for motion, pacing or editing rhythm. Cite as Video 1.
Up to 3 audio tracks (2-15s each, combined up to 15s) to sync motion, voice or music. Needs at least one image or video reference.
Duration
Length of the generated video in seconds (5-15). Billed per second.
Resolution
480P and 768P are native; 2K and 4K upscale a 768P base render.
Aspect Ratio
Aspect ratio. 'adaptive' follows the reference media.
50% less than the priciest provider (4 compared). Same model, same output.
Your result renders here.
Use MiniMax H3 Reference to Video via the API
Generate with MiniMax H3 Reference to Video from your own code. Async by design: submit a generation, get a run id back instantly, poll until it completes. Billed in credits at the best available provider price — from $0.04 per second, audio included.
1 — Start a generation
curl -X POST https://app.scenetra.com/api/v1/generate/minimax-h3-ref2v \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"prompt":"Image 1 walks through Image 2 at golden hour, handheld follow shot, crowd murmur","reference_image_urls":["https://example.com/person.jpg","https://example.com/street.jpg"],"duration":"5"}'Response (202)
{
"runId": "run_abc123",
"status": "PENDING",
"model": "minimax-h3-ref2v",
"estimatedCredits": 650,
"statusUrl": "/api/v1/runs/run_abc123"
}2 — Poll until COMPLETED (every 2-5 seconds)
curl https://app.scenetra.com/api/v1/runs/run_abc123 \
-H "Authorization: Bearer YOUR_API_KEY"Response
{
"runId": "run_abc123",
"status": "COMPLETED",
"progress": { "completed": 1, "total": 1 },
"outputs": [
{ "type": "video", "url": "https://images.scenetra.com/..." }
],
"error": null,
"estimatedCredits": 650
}Request fields
| Field | Type | Description | Default |
|---|---|---|---|
| prompt* | string | Describe the result and cite reference inputs in order: Image 1, Video 1, Audio 1, and so on. | — |
| reference_image_urls | string[] (https URLs) | Up to 9 reference images for subjects, products, style or composition. Cite in the prompt as Image 1, Image 2. | — |
| reference_video_urls | string[] (https URLs) | Up to 3 reference clips (2-15s each, combined up to 15s) for motion, pacing or editing rhythm. Cite as Video 1. | — |
| reference_audio_urls | string[] (https URLs) | Up to 3 audio tracks (2-15s each, combined up to 15s) to sync motion, voice or music. Needs at least one image or video reference. | — |
| duration | string | Length of the generated video in seconds (5-15). Billed per second.5 · 6 · 7 · 8 · 9 · 10 · 11 · 12 · 13 · 14 · 15 | 5 |
| resolution | string | 480P and 768P are native; 2K and 4K upscale a 768P base render.480P · 768P · 2K · 4K | 2K |
| aspect_ratio | string | Aspect ratio. 'adaptive' follows the reference media.adaptive · 21:9 · 16:9 · 4:3 · 1:1 · 3:4 · 9:16 | adaptive |
| callbackUrl | string (https URL) | Optional. When the run finishes (completed or failed), we POST the final run status to this URL — same shape as the poll response, plus model, params, and chargedCredits. Best-effort; polling remains the source of truth. | |
Unknown fields are rejected with a helpful 400. Media fields take https URLs. There is no provider parameter — every run is routed to the cheapest capable provider automatically.
Good to know
Billing
Runs bill your account credits at provider cost with 0% markup. The submit response includes the credit estimate; failed generations are refunded automatically. Add ?dryRun=1 to get the estimate without running anything.
Errors
Errors are always { "error": { "code", "message" } } with stable codes: unauthorized, model_not_available, invalid_parameter, insufficient_credits, rate_limited. A failed run is not an HTTP error — poll responses report status: "FAILED" with the reason.
What makes MiniMax H3 Reference to Video different
Audio is part of the generation, not an add-on
H3 renders stereo audio in the same pass as the picture: spoken dialogue, background music, and sound effects that match what's on screen. Write the sound into your prompt the way you'd write the visuals — a line of dialogue, a music cue, a door slam — and it arrives synchronized, with no separate audio pass and no surcharge on the per-second rate.
Four resolution rungs, from drafts to 4K
480P and 768P are native renders; 2K and 4K upscale a 768P base for delivery-ready output. Unusually, 768P is the cheapest rung at $0.04 per second — below even 480P — which makes it the natural default for iteration. When a take lands, flip the same node to 2K ($0.065/s) or 4K ($0.16/s) and re-run.
5 to 15 seconds, any whole second
Duration is a per-second choice from 5 to 15, not a couple of fixed presets. Five seconds is a cheap motion test; fifteen is enough for a full beat with dialogue and a music cue. Because billing is per second, a 7-second clip costs exactly 7 seconds — you never pay for a 10-second slot you didn't use.
Three modes on one board
Text to Video is one of three MiniMax H3 modes on Scenetra. Animate a first-frame image (optionally pinning the last frame) with Image to Video, or drive generations from up to 9 reference images, 3 reference clips, and 3 audio tracks with Reference to Video. All three run as nodes in the same visual workflow, so you can chain an image model straight into H3.
Parameters
| Parameter | Type | Description | Default |
|---|---|---|---|
| Prompt* | text | Describe the scene, action, camera movement, visual style, dialogue, music, and sound effects — H3 renders audio natively. | — |
| Duration | select | Duration of the video in seconds (5-15).5 · 6 · 7 · 8 · 9 · 10 · 11 · 12 · 13 · 14 · 15 | 5 |
| Resolution | select | 480P and 768P are native; 2K and 4K upscale a 768P base render.480P · 768P · 2K · 4K | 2K |
| Aspect Ratio | select | Aspect ratio of the generated video.21:9 · 16:9 · 4:3 · 1:1 · 3:4 · 9:16 | 16:9 |
Pricing
From $0.04 per second, audio included
Pay per generation — only pay for what you use.
480P
$0.05/s
per second
768P
$0.04/s
per second
2K
$0.065/s
per second
4K
$0.16/s
per second
What a video costs
| Duration | 480P | 768P | 2K | 4K |
|---|---|---|---|---|
| 5 seconds | $0.25 | $0.20 | $0.33 | $0.80 |
| 10 seconds | $0.50 | $0.40 | $0.65 | $1.60 |
Use cases
Film & storytelling
- Dialogue scenes with synced voices
- Establishing shots in 21:9
- Mood pieces with score and ambience
- Previz with temp sound built in
Marketing & ads
- Product spots with music and VO
- UGC-style ads with natural dialogue
- Brand stings with sound design
- A/B variants at draft resolution
Social content
- Vertical 9:16 clips with audio hooks
- Talking-character shorts
- Sound-on meme clips
- Music-driven transitions
Music & performance
- Performance clips with generated music
- Rhythm-matched visuals
- Lyric-video backgrounds
- Ambient loops with soundscapes
Frequently asked questions
What is MiniMax H3 Reference to Video?+
It's the H3 mode that generates video from reference material instead of just a prompt: up to 9 reference images, 3 reference video clips, and 3 audio tracks in one generation. Images anchor subjects and style, videos contribute motion and pacing, and audio drives timing — while H3 still renders native audio in the output. On Scenetra it runs as a node in the visual workflow editor.
How do I reference the uploads in my prompt?+
By slot, in order: Image 1, Image 2, Video 1, Audio 1, and so on. Write the prompt as direction — 'the woman from Image 1 walks through the cafe from Image 3, moving to the rhythm of Audio 1' — so the model knows which reference plays which role in the scene.
What are the limits on reference videos and audio?+
Up to 3 reference video clips, each 2-15 seconds, with a combined length of 15 seconds or less — and the same limits for audio: up to 3 tracks, 2-15 seconds each, 15 seconds combined. Audio references also need at least one image or video reference alongside them in the same generation.
Can I keep the same character across multiple videos?+
Yes — that's the core use case. Put the character's reference images on a Scenetra board, wire them into several H3 Reference to Video nodes, and change only the prompt per shot. Because every generation sees the same references, the face, outfit, and styling stay consistent across the whole sequence.
Related models
Happy Horse 1.1
Alibaba's Happy Horse 1.1 generates 3-15 second videos with native audio straight from a text prompt. Dialogue, music, and sound effects are directed through the same prompt as the visuals, and nine aspect ratios from 9:21 to 21:9 cover everything from vertical stories to ultrawide cinematics at 480p, 720p, or 1080p.
View model →Seedance 2.5
ByteDance's Seedance 2.5 generates cinematic videos up to 30 seconds from text prompts with synchronized audio — voice, sound effects, and background music — built in by default. The newest Seedance generation, with mp4 or high-fidelity mov output for multi-round editing pipelines.
View model →Kling 3.0 Pro
Kuaishou's flagship video generation model delivering stunning visual fidelity. Kling 3.0 Pro supports text-to-video, image-to-video with start/end frames, element references for character consistency, and native audio generation.
View model →Grok Imagine Video 1.5
xAI's Grok Imagine Video 1.5 turns a text prompt into a video with synchronized audio generated in the same pass — no separate sound step and no audio surcharge. Choose any duration from 1 to 15 seconds, three resolutions from 480p to 1080p, and eight aspect ratios from widescreen 16:9 to vertical 9:16.
View model →Start creating with MiniMax H3 Reference to Video
Use MiniMax H3 Reference to Video alongside 50+ other AI models in Scenetra's visual workflow editor. No setup required.
Get Started