Scenario Minimax Video skill

Scenario MiniMax Video

Overview

MiniMax's Hailuo video family on Scenario has three kinds of member. H3 (Hailuo 3.0) folds text, keyframe, and reference conditioning into one model and generates stereo audio in the same pass. The H3 Max line splits that into single-mode members (text, image, reference, lip sync), each with a faster Turbo twin on the first two. Three Max members edit a clip you already have instead of generating one: Extend Video (with a Turbo twin), Insert Video, and Recast. The older Hailuo 2.3 pair carried a deprecated tag naming H3 as successor at authoring time. Discover members with search and treat model_schema_get as the contract: they agree on almost nothing, not even the spelling of a resolution.

Connection and the core loop: see the scenario skill in this repo; model-agnostic video work: the scenario-video skill. If a sibling skill named here is missing from your available skills, load it with skill_get (Scenario MCP) or ask the user to install it (https://github.com/scenario-labs/skills#install); unattended, proceed from tool schemas and flag the gap.

Quick reference

H3's mode follows from the inputs (names from the live schema):

Mode	Inputs	Behavior
Text	prompt	aspectRatio honored (21:9 through 9:16, default adaptive)
First frame	firstFrameImage (+ prompt)	shape follows the image; aspectRatio ignored
First + last	firstFrameImage + lastFrameImage	lastFrameImage is valid only alongside firstFrameImage
Reference	referenceImages, referenceVideos, referenceAudio	guides subject, style, motion, and voice, not forced frames; aspectRatio still applies

Keyframes and references are mutually exclusive: neither frame combines with any reference array. referenceAudio never rides alone; it requires at least one image or video reference. Reference parameters are arrays even for one asset. At authoring time H3 took 9 reference images, 3 videos, and 3 audio files (videos and audio each 2 to 15 seconds, and 2 to 15 seconds in total), 5 to 15 seconds of output, at 768P or 2K.

The 2.3 members take only prompt and firstFrameImage, plus a coupled pair: at authoring time 10 second duration was available only at 768p, and 1080p only at 6 seconds. Their promptOptimizer (default true) rewrites the prompt before generation: leave it on for thin prompts, switch it off when engineered wording must survive verbatim. H3 has no such switch, and neither H3 nor 2.3 takes a seed.

Picking the member

H3 is the one member that mixes modes in a single call, and, lip sync aside, the one generator that reaches 2K; it is also slow and expensive, and reference media adds more (reference videos bill per second of uploaded footage). The Max text and image members cap at 768P and take a seed and a promptExpansionMode (disabled, balanced, quality); set it to disabled when engineered wording must survive verbatim. Their Turbo twins share the same inputs and are the iteration tier: draft at 480P on Turbo, then spend the expensive member on the take you keep. A draft is a composition and motion check, not a preview of the final pixels: a different member or resolution is a new generation even with the same seed. dry_run the draft and the final payloads before a batch. Re-discover a member before naming 2.3 for new work: deprecated members can disappear from the catalog.

Editing a clip you already have

These members take the existing video as video (any video asset id, generated or uploaded) and bill only what they make or touch, so price each with dry_run on the real source.

Member	What it does	Key inputs and caps at authoring time
Extend (+Turbo)	adds 1 to 15 s after the source	source 1.625 to 60 s, up to 50 MB, aspect 0.4 to 2.5; output extended returns source plus new footage, continuation the new footage behind a lead-in copied from the source (see Common mistakes); referenceAudio replaces the source soundtrack as the audio guide; aspectRatio other than auto crops; 480P to 2K
Insert Video	a new 5 to 13 s scene, then the original resumes	startTime and resumeTime in source seconds (both required); duration is the new scene only; up to 9 referenceImages and 3 referenceVideos, billed by reference tokens; colorMatch on by default; 480p or 768p, lowercase
Recast	swaps up to 4 people for new ones from photos	source 5 to 30 s, no single shot over 15 s, billed per source second; referenceImages one photo per person, required, mapped left to right by default; prompt optional, to say who becomes whom; motion, camera, cuts, and audio are kept; 768P or 1080P

Write an Extend prompt about what happens next, never a description of the source: the model already sees it, and enablePromptExpansion (on by default) rewrites the prompt from the source to keep the continuation consistent. Recast is the lane for re-shooting the same performance with a different person; it keeps everything else, so it is the wrong tool for a new setting or style.

Camera in brackets, motion in moderation

The whole family reads bracketed camera commands inline in the prompt: [Push in], [Pull out], [Pan left], [Tilt up], [Truck right], [Pedestal up], [Zoom out], [Shake], [Tracking shot], [Static shot]. Up to three moves combine in one bracket ([Pan left, Pedestal up]); separate brackets sequence them. Keep 2 or 3 motion cues per shot in total: piling on moves invites background wobble and texture flicker. Image-to-video tends to drift even unprompted, so lock statics explicitly ("Static camera, locked shot, tripod mounted").

Write the rest as natural prose, ordered camera, subject, action, scene, lighting and mood, style. On H3, sound lives in the prompt too: describe dialogue lines, SFX, and ambience inline so they sync to the action; no audio parameter exists to switch instead.

Worked example: a character clip from reference stills

1. search with target="models", query="minimax hailuo", public=true. Prefer the newest non-deprecated hit, e.g. model_minimax-h3 (a live hit at authoring time: re-discover each session).
2. model_schema_get with that id: modes, caps, and allowed values before anything else.
3. upload_asset two clean, evenly lit character stills (see the scenario skill) to get asset ids.
4. model_run with that model_id, dry_run=true, and the exact parameters={"prompt": "[Tracking shot] The scout from the reference images sprints across a rooftop at dusk, coat snapping in the wind, warm rim light, footsteps and distant traffic in the audio, cinematic realism.", "referenceImages": ["asset_a", "asset_b"], "duration": 8, "resolution": "2K", "aspectRatio": "16:9"} for the cost estimate; re-estimate after changing duration, resolution, or the reference count.
5. Repeat model_run with wait=false, then jobs_wait with the returned job id, re-called with pending_job_ids on timeout, never a second model_run; H3 at 2K commonly runs several minutes.
6. asset_display the output and review it with sound on: dialogue and SFX sync are part of what you paid for.

Common mistakes

• Combining firstFrameImage with a reference array: keyframes and references never mix.
• Passing referenceAudio alone: it requires at least one image or video reference.
• Sending lastFrameImage without firstFrameImage: the pair anchors both endpoints or neither.
• Expecting aspectRatio to win over a first frame: the shape follows the image.
• Carrying one member's values to another: H3 spells resolutions 768P and 2K, the Max generators other than lip sync stop at 768P, Insert spells 480p and 768p in lowercase, and Recast offers only 768P and 1080P.
• Sending Recast a clip with a single shot over 15 seconds, or one under 5: cut it first (the scenario-video-editing skill).
• Pricing Recast on the output: it bills every second of the source, so trim the source to the part that needs the new people.
• Expecting either Extend output to be new footage only: extended returns the source plus the continuation, and continuation opens on a copy of the source's last 1.625 s (39 frames at 24 fps in two authoring-time runs), then new footage that ran a little past the request (2 s gave 2.1 s, 3 s gave 3.5 s, each billed as requested). For pieces assembled later (the scenario-video-assembly skill), take continuation and cut from startTime 1.7: the cut tool steps by 0.1 s and 1.6 keeps a source frame (the scenario-video-editing skill).
• Stacking camera moves: past 2 or 3 cues the background wobbles and textures flicker.