Video Model Lab

Work protocol

One source, one brief, many models. No promotion without repeatable evidence.

Lab rule

Production stays calm. The Lab gets curious.

Traveler can keep using the current private pipeline while this room tests better video options. The goal is not to chase novelty every day. The goal is to notice when a model becomes meaningfully safer for S.

Use the same Traveler source image across all candidates in a run.
Use the same city, motion brief, duration target, and safety constraints.
Score the output before knowing which model made it whenever possible.
Record funny failures because they are teaching material.
Only promote a model after multiple clean private tests, not one lucky render.

First incident

The magic bag test

The latest private Traveler Short produced a useful hallucination: S's bag appeared carried by an invisible hand. That failure is now part of the rubric, not just a joke.

Known failure mode Object contact must remain visible. If S carries a bag, suitcase, phone, cup, umbrella, or prop, her hand must stay physically connected to it throughout the shot.

Model bench

Candidate slots are intentionally provider-neutral. Availability and cost can change quickly, so this page tracks the workflow, not a permanent ranking.

Slot	Role	What we test	Promotion signal
Seedance	Current baseline	Long uninterrupted motion, Traveler cadence, current prompt guards.	Keep if it remains most reliable after object-contact and signage tightening.
Sora candidate	High-priority challenger	Realism, motion physics, identity preservation, natural camera language.	Promote only if it beats baseline in repeated private Traveler shots.
Veo / Runway / Kling / Pika	Rotating challengers	Continuity, hands, bags, signage, face drift, duration, cost, speed.	Earns a production trial by winning a controlled bakeoff, not by hype.
No-video fallback	Safety valve	Ken Burns stills, parallax, slow editorial motion, image sequence cuts.	Use when model video risks identity or physical weirdness on a public day.

Rubric

Score each candidate 1-5. Anything below 4 on identity, physics, or public safety cannot graduate.

S identityFace, age, body language, editorial presence, and recognizability.

PhysicsHands, bags, walking, object contact, cloth, hair, and gravity.

City truthSevilla should feel like Spain, not generic Europe or wrong-language signage.

ContinuityNo drifting face, sudden outfit changes, warped props, or broken scene logic.

Short fitDuration, composition, pacing, readable thumbnail, and safe crop for Shorts.

Teaching valueWhat this output teaches AB about current AI capability and limits.

Prompt guard

This is the first shared safety language for object contact and localized realism.

Reusable motion constraint

If S carries any bag, suitcase, phone, cup, umbrella, or object, her hand must remain visibly in contact with it throughout the entire shot. No floating objects, invisible support, detached accessories, extra hands, or object drift. Keep S moving slowly and naturally. Prefer subtle camera movement, wind, light, and environmental motion over complex hand-object choreography. All visible signs, menus, street text, and written language must match the location. For Sevilla, use Spanish-language public text only, or avoid readable text entirely.

Linked rooms

Use these together

Video testing depends on stable source images and a clean production gate.

S Consistency Lab source images Production Pipeline runtime Traveler Room canon

Promotion gate

The Lab can recommend; production only changes after AB approval.

Stage 1

Collect

Save the source image, model, prompt, cost, duration, and output URL for every serious test.

Stage 2

Score

Use the six-part rubric and write down the failure mode, even when the result is charming.

Stage 3

Private

Run at least three private Traveler uploads with the candidate before any public schedule change.

Stage 4

Promote

Switch production only when the new model is better, safer, and easier to explain than the current baseline.