A multi-shot video is easier to review when each shot has one purpose and a clear acceptance rule. This tutorial uses minimax-h3 to turn a short product reveal into a shot list. It does not promise that separately generated clips will share identical geometry, lighting or motion.
Plan the edit before generating clips
Section titled “Plan the edit before generating clips”Write the sequence in an editor’s terms. For a desk-lamp reveal, use the following plan; durations are chosen examples, not measured model performance.
| Shot | Intended action | Continuity anchor | Review gate |
|---|---|---|---|
| A: introduction, 5 seconds | Fixed view of the lamp on a desk | Lamp shape and desk position | No shape change or extra objects |
| B: detail, 5 seconds | Slow move toward the lamp shade | Same material and color | Shade remains recognizable throughout |
| C: closing, 5 seconds | Hold the final composition | Same orientation and lighting | Stable closing frame for later titles |
Keep exact brand marks and final captions as separate editing assets. A generated shot is not a reliable place to enforce precise typography or legal copy. The shot plan is local production metadata, not an extra field in the API request.
Pick a reference mode for each shot
Section titled “Pick a reference mode for each shot”Use frames_images when the opening frame, or opening and closing frames, should guide a shot. One URL supplies a start frame; two URLs are ordered start then end. Use H3’s general images references when you want subject guidance without defining the opening frame. Do not mix frames_images with images, videos or audios.
Use authorized assets and retain their versions. For frame mode, prepare the reference crop deliberately; the generated composition may follow the source frame rather than a conflicting aspect-ratio instruction. Check the H3 / Turbo selection guide for resolution and mode constraints before choosing a route.
Submit a single planned shot
Section titled “Submit a single planned shot”This request body describes shot B with an opening frame. Replace the placeholder URL with a reachable approved image and verify the full live schema before submission.
{ "model": "minimax-h3", "input": { "prompt": "One continuous shot of the desk lamp in the opening frame. The camera moves slowly toward the shade. Preserve the lamp silhouette, cream finish and desk position. Soft steady light. No cuts, no extra objects, no lettering.", "frames_images": ["https://assets.example.com/lamp-shot-b-start.png"], "seconds": 5, "resolution": "768P", "aspect_ratio": "16:9" }}Send it to POST /v1/tasks using the H3 API guide. Save the returned ID against shot B in your own shot manifest. Follow the Node.js task workflow to resume polling that ID. Three shots are three separately managed tasks; this guide does not imply a storyboard or batch-submission endpoint.
Review the whole clip and each join
Section titled “Review the whole clip and each join”Watch each complete clip, not only its first frame. Check object persistence, camera direction, motion, frame edges and any audio you intend to use. Then place adjacent clips in the edit and inspect the cut: subject scale, position, color temperature and movement should serve the planned transition.
If you want to use a frame extracted from an accepted clip as the next opening reference, extract and approve it in your own editing workflow. The next request is still a new paid task and may drift. Keep the original product reference for comparison rather than treating generated frames as unquestionable ground truth.
Recover at the shot level
Section titled “Recover at the shot level”If a task fails technically, record its error and charge outcome. If it succeeds but misses the creative brief, mark the shot rejected and revise only that shot’s prompt or reference. A successful-but-unusable clip is not fixed by automatically retrying the same network request.
Budget for attempts per shot and review cost per accepted sequence using batch budgeting. Keep shot ID → brief version → reference version → task ID → review result → charge together. Do not regenerate already accepted shots merely because another shot needs correction.
Contract and source check: 2026-11-03. MiniMax’s native video API uses a different request shape; the platform task uses model and input as shown above. No generation was performed to establish this workflow, and it makes no measured continuity claim.