Scenario Consistency

Carrying identity into video

Reference-to-video members animate the subject the references show, with no first frame locking composition. The catalog has no reference-to-video capability value: these members are tagged img2video, so recommend with capability="img2video" and the consistency need in the user's own words ("the same character as in these images"), then confirm on model_schema_get that the pick carries a multi-slot referenceImages array (7 to 9 slots on the leading members at authoring time, one of them billing slots past the first four) and, to copy a motion, a referenceVideo or referenceVideos slot (capped between 3 and 15 seconds, billed per second of upload on some). When nothing ranked carries the array, search a family name from the video model-family skills scenario-video lists. Slot count moves the price less than resolution and duration: read both enums off model_schema_get, where cost_impact marks them, and prove identity on the cheapest roll that still shows the whole action, the lowest resolution at the delivery duration (a shorter clip hides late drift), pricing that payload and the delivery payload with dry_run first (one proving roll priced 80 CU against 130 at the defaults). A passing proving roll validates the recipe, not the roll: the delivery render gets its own sweep before it ships. image, the first-frame anchor, and the reference arrays are mutually exclusive on several families: a reference run opens free, and pinning the opening frame is a different member. Fill the slots from the library: the hero and two or three approved on-model shots per scenario-identity-library, as individual shots, never the turnaround sheet. With no library yet, the approved hero alone in one slot is a valid first run (the published pipeline starts from one strong reference), and the next slots fill from approved stills out of the image reference loop above, once one reads as a clean character view (neutral pose, no scene props: a scene still carries its props and pose into the slot), never from cropping the sheet. Prompt the action, not the look: some schemas address uploads by order ("Image 1 turns toward the camera, moving as in Video 1"), others state that the subjects appear with or without a prompt, and on all of them re-describing the character fights the images. Per-family contracts and prompt syntax: the video model-family skills scenario-video lists. Judge identity across the whole clip: asset_get returns firstFrame and lastFrame as free asset ids, but a drift that fades in mid-shot passes both, so sweep the frames with model_scenario-video-to-image-seq (a fixed first-party id: Scenario's single deterministic frame extractor, so discovery would only re-derive it) before the clip ships: about two frames a second (frameInterval is an integer count of source frames between extractions, the clip's frameRate from asset_get halved and rounded, 12 at 24 fps; the clip goes in video), since a spot check of one frame per shot read clean at 10.5 seconds on a clip whose defect appeared at 11.0. Judge each frame against the baseline enumeration by sub-element: a named part gained, lost, recolored or reshaped is drift; a part that moves with the action, a cloak swinging or an antenna tilting, is animation and passes.

When references stop scaling

Reference-first is the published order and the cheaper one, so exhaust it before training. Training pays back when one character recurs across hundreds of images, when four or more identities must share scenes, or when every run needs the same rendering style whatever the subject. Train one model per recurring character and a separate one for the style: a single model carrying both drifts on whichever it saw less of. Multi-character scenes composite rather than co-generate, because reference slots split attention: generate each identity alone against its baseline, then place them together with the edit workflows in scenario-image-editing. Dataset size, base choice, and the quote-before-launch rule: scenario-model-training. A trained model never runs by its own id (runs_as and run_with, see scenario), and it still takes the baseline-plus-delta prompt: training lightens the enumeration, not the delta discipline.

Sprite sheets and animation frames are the hard case: adjacent frames from a general image model drift in proportion, scale, and silhouette, which an engine exposes at once. Frames of one character come from the video lanes in scenario-sprite-animation, where one render holds the identity instead of a generation per frame.

Worked example: five poses of one mascot

1. asset_display the approved hero (asset_hero) and write its baseline: the full must-not-change enumeration above.
2. recommend with the task's own words as prompt (search only for a named family), preferring models with reference-image slots, then model_schema_get: the reference field's name, cap, cardinality, requiredness. No reference field in the schema disqualifies the candidate: go back to the ranked list, and when that holds nothing reference-capable either (recommend ranks community fine-tunes and can miss first-party models), take a family name from a sibling model-family skill (scenario-gemini-image, scenario-seedream) and search for it.
3. One model_run per pose, five in all, launched with wait=false in waves within the concurrency ceiling: the byte-identical baseline, the pose alone in the final clause, the hero in the reference field shaped as the schema says: ["asset_hero"] only under array: true. No seed. No control map: a pose map from the hero locks the pose being changed.
4. jobs_wait on each wave, re-calling with pending_job_ids, and launch the next pose as a slot frees until all five are done. asset_display each against the hero; fix drift by tightening the enumeration, not by chaining outputs.

Common mistakes

• Reaching for a seed to make two prompts match: it reproduces one generation and transfers nothing. Leave it unset across a set; set one only to re-roll a single unchanged prompt.
• Writing "same as before": there is no memory between calls; restate the baseline in full each time.
• Chaining a set output to output: drift compounds. Anchor every item to the same approved baseline.
• Training for a one-off: one character in one scene is a reference job, and a recurring cast at scale is the training case above, one model per character plus one for the style.
• Assuming asset_detect modality names are valid controlModality values.
• Passing the turnaround sheet into a reference slot, video or image: a video member animates the grid, and on an image model the sheet is a filing and review artifact, not a reference. Slots take individual approved shots.
• Re-describing the character in a reference-to-video prompt: the images carry the look; the prompt carries action, camera, and timing.