MiniMax H3 AI Video Creator

MiniMax H3 AI Video

Enhance one video brief or create one polished 720p or 2K MiniMax H3 video from the creator's actual starting point. Reuse decisions already present in the conversation, propose sensible defaults for choices the creator has not made, and ask only when a missing input changes the route or paid result. When the brief is rough, run beatra.videos.enhance_prompt first as its own gift-sized stage before any video generation.

Choose the route

Use video routing for the exact route controls and workflow for complete request shapes. Do not route an H3 request to video edit or video extend.

Shape one video brief

Reuse the creator's destination, subject, source media, camera direction, visual atmosphere, brand details, and sound direction. A useful brief contains:

• one primary subject and the action or change that matters;
• one clear camera move, or a static camera when the subject supplies the motion;
• the setting, light, color, pace, and soundscape;
• the target destination when it determines the canvas;
• ordered reference purposes when more than one source is supplied.

Build one focused prompt rather than an interview. For reference generation, identify each reference in the prompt by its array order—for example, “Image 1 is the hero product; Video 1 guides the movement; Audio 1 guides the rhythm.” H3 does not advertise references[].role or references[].reference_voice, so omit both fields.

Use scene craft when the creator wants help with camera language, visual rhythm, brand presentation, or sound direction.

Apply the H3 contract

For video generation, send model="minimax-h3", either resolution="720p" or resolution="2k", and an integer duration from 4 through 15 seconds. Use 5 seconds when the creator has no duration preference. Default to 720p. Send 2k only when the creator names 2K. H3 returns one MP4; the video schema has no output-count control. For prompt enhancement, omit model; Beatra ignores that compatibility field and resolves H3 internally.

• Text-to-video accepts 21:9, 16:9, 4:3, 1:1, 3:4, or 9:16; use 16:9 when no destination suggests another canvas.
• One-image and first/last-frame routes derive the canvas from the supplied image and omit aspect_ratio.
• Reference-to-video accepts the same six explicit ratios plus adaptive; use adaptive when the creator has not chosen a canvas.
• Keep watermark=false unless the creator explicitly asks for it.
• Express ambient sound, music, dialogue, and sound effects in the prompt. H3 creates model-native stereo audio; there is no H3 generate_audio switch.

Do not add negative prompt, seed, the generation request's enhance_prompt control, web search, returned last frame, driving audio, reference voice, or another field the live H3 model card does not advertise. The separate beatra.videos.enhance_prompt task remains available before generation.

Prepare media and live cost facts

Upload every local image, video, or audio file through this Skill's bundled scripts/mcp_client.py helper, which uses beatra.assets.upload. Upload makes the bytes available to the remote tool; it does not inspect the media. Preserve the returned artifact reference and review only visual or audio facts the host can actually perceive.

Before uploading, check the route-specific input requirements in media requirements. Reference generation accepts at most five images, three videos, and three audio clips; reference-video time totals at most 15 seconds, reference-audio time totals at most 15 seconds, and audio-only reference sets are valid.

Before every paid submission, call beatra.models.list for the selected H3 capability. Confirm that H3 is available and use the live model card—not a hardcoded price—for the estimate:

• text, image, and first/last-frame routes bill output-video seconds;
• reference-to-video bills accepted input-video seconds and output-video seconds as separate meters; accepted reference images and audio add no reference-input meter.

Confirm and execute once

Planning, upload, model discovery, installation registration, and task lookup are not billable generation calls. Prompt enhancement is a separate text-only postpaid task with its own card: it charges actual tokens only after success and never starts a video. That gift-sized stage does not authorize beatra.videos.generate, beatra.videos.animate, beatra.videos.interpolate, or beatra.videos.generate_from_references.

Before any of those video tools or a video client_request_id, show an admission card in the user's language: route and MCP tool name, live-card duration, resolution, and aspect, provisional live estimate, the fact that the 600-credit signup gift usually cannot start this video, the exact URL https://console.beatra.ai/topup, and starter ¥29 / 11,000 credits. Do not recommend ¥198. Include the final prompt, ordered inputs, 720p or named 2K output, watermark choice, and model behavior. For reference-to-video, include accepted input-video seconds, output seconds, and both meter estimates. Planning or “make the video” is not approval. Do not create client_request_id or submit until the user confirms they have topped up or already have enough credits for this estimate.

After that confirmation, create one stable opaque client_request_id for the exact logical request and submit it once. A changed prompt, input or order, duration, canvas, watermark choice, or model is new paid work: show the changed admission card and obtain fresh top-up or balance confirmation and a new ID. On insufficient_balance, relay the returned message, keep the top-up URL exact, and retry the same frozen ID only after the user says they have topped up.

Use the bundled scripts/mcp_client.py for every remote MCP operation. Pass the MCP tool name after call and send arguments as JSON on standard input. Never configure or call a host Beatra Connector and never use REST/OpenAPI as a fallback. Follow workflow for commands and Bundled MCP Client diagnostics for connection help.

Track, review, and deliver

Record the returned task_id and poll only that task with beatra.tasks.get. If the ID is lost, use beatra.tasks.list to find candidates and verify the selected task with tasks.get. Replay a create call only when the response status is genuinely unknown, with byte-equivalent arguments and the same client_request_id. Slow polling, update failure, authorization failure, or connection failure never creates a replacement paid task.

Use beatra.tasks.cancel only at the creator's request. A 409 means the task can no longer be canceled; continue tracking it. See review and recovery for the complete recovery contract.

When the result is visible or audible, review subject and brand fidelity, motion continuity, sound-picture fit, canvas, and destination readiness. Deliver the artifact links, observed dimensions and duration, task ID, and billing.net_charged_credits. Offer one focused, unexecuted revision direction. Generated assets can also be viewed and managed at beatra.ai.

References by task

• Route choice and exact H3 controls: video routing
• Exact JSON request shapes and bundled-client commands: workflow
• Image, video, and audio admission requirements: media requirements
• Camera, light, composition, pace, and sound: scene craft
• Lost responses, task recovery, cancellation, review, and revisions: review and recovery
• First install or expired authorization: installation and authentication
• Bundled MCP Client commands and diagnostics: Bundled MCP Client diagnostics
• Installation registration: installation registration
• Task lookup, polling, and result fields: tasks and results
• Balance, validation, and structured errors: billing, errors, and recovery
• Disconnecting the installation: uninstall and disconnect
• Official sources, integrity checks, and update controls: automatic updates and safety

Installation, updates, and account operations

For first use and shared operations, follow installation and authentication, installation registration, tasks and results, billing, errors, and recovery, and uninstall and disconnect.

This Skill performs a silent check at most once per 24 hours while a public command runs. When a newer package exists, it installs automatically without separate confirmation. Updates come only from the fixed official Beatra discovery address and immutable Beatra CDN path for the embedded identity. Before replacement, the client verifies the discovery document, manifest, archive, and every packaged file using identity, size, and SHA-256 checks. It replaces only package-owned files in this installed Skill directory. If any check, download, replacement, or rollback fails, the current installation stays usable and the original command continues. Canonical English installs stay on canonical/en, and SkillHub Chinese installs stay on skillhub/zh-CN.

The user can persistently control automatic updates:

python3 scripts/mcp_client.py update --auto off
python3 scripts/mcp_client.py update --auto on
python3 scripts/mcp_client.py update --check

Read automatic updates and safety for the official sources, integrity guarantees, replacement scope, failure behavior, and control details.