Video Model Selection

Mission

Turn a creative video requirement into an evidence-backed model/endpoint decision. Treat model catalogs, wrapper repositories, and remembered names as leads; the live schema, availability, cost estimate, and returned output shape control execution.

Use Video Strategist when selection is one part of a larger brief. Use this skill for a direct “which model/path should I use?” request or before a material batch.

Required inputs

• text-to-video, image-to-video, reference-video, edit, extend, motion, lipsync, clip, caption, or enhancement intent;
• source/reference count and roles;
• target duration, ratio, resolution, frame rate, audio, language, and motion;
• continuity and product/likeness invariants;
• latency, quality, budget, batch size, and retry tolerance; and
• connected tools, account/region constraints, and rights state.

Missing requirements must remain unknown. Do not rank a model by marketing claims, a stale README, or an editorial “virality” score.

Selection procedure

1. Translate the brief into hard constraints and preferences. Separate requirements that belong to generation from those better handled by editing, captions, audio, or local post-production.
2. Use media.search_models when available, then verify the exact endpoint, request schema, supported reference fields, duration/ratio choices, output shape, retention, and current cost. See MODELS.md for a dated orientation only.
3. Eliminate candidates that cannot satisfy hard constraints. Record why a model is unsupported, unknown, too expensive, or semantically wrong; do not silently substitute T2V for an edit or I2V for a multi-reference request.
4. Compare the remaining candidates on continuity, camera/motion control, native audio, text/label risk, latency, quality, failure modes, and total call count. Check balance before a material batch.
5. Recommend one primary, one fallback, and a proof render. State fields that still need live validation and the stop condition for retries.

Output contract

Return a requirements table, live discovery timestamp, candidate endpoints and schemas, eliminated options with reasons, price/credit evidence, primary and fallback choice, proof-render plan, approval needed, and expected QA risks. Do not claim that a model is best in general or that a proof render guarantees the final sequence.

Run video QA on the proof render and use muapi-video-tools.md for current connection and lifecycle details. This skill recommends a path; it does not submit a paid batch without approval.