Scenario Minimax Video

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.