"Make variant two look exactly like variant one except for X" is the most repeated creative ask, and agents reach for seeds, which do not solve it. Consistency comes from what you feed the model, in rising order of durability: a prompt baseline, reference images, a style reference, a control map, a trained model. The same references carry a character into video. Connection and the core loop: see the skill; training: see ; creating and filing a named character or prop library, with its sheets: see . If a sibling skill named here is missing from your available skills, ask the user to install it (); unattended, proceed from tool schemas and flag the gap.

asset_detect

Scenario's published pipeline guidance: generate one strong reference image, pass it as an image input to every scene generation, and prefer models with the most reference-image slots.

Write out everything that must not change, then put the single change in a final clause. Keep the prompt byte-identical between runs and edit only that clause. A change every item must carry (new wardrobe, a haircut) is baseline, not delta: run it once as the hero's own delta, approve the result as the new hero, then anchor the set to it, so the reference and the enumeration describe the same look.

Enumerate specifics: subject geometry, camera height and angle, subject size and position in frame, lighting direction, each named sub-element and where it sits, palette by name or hex, and embedded text. Look at the baseline first (): you cannot enumerate a shade you have not seen, and vague anchors drift.

Attach the approved baseline as a reference image alongside it. Image models converge on , but cap, requiredness, and cardinality come from , and on some the field is a single scalar file. A scalar plus is img2img, not a reference slot: it anchors composition along with identity, and at the default strength a plain-background anchor overrides a whole-scene delta, so a scene set needs a true reference field. Wrap an array only where the schema says , even a lone asset as : a bare string is silently dropped and the run succeeds ignoring it. State each reference's role in the prompt. Approved on-model shots fill the remaining slots: curate them as a collection (catalog tools, write lane) and retrieve them with . Some style-reference members mint a reusable style on the first run: the id sits in the finished job's output metadata ( with ), not in the run response or on the asset, and later runs pass it in the schema's id field in place of the images. The two fields are exclusive and only the image field is flagged , so the mint is paid once: read the field text before passing the same reference set on every run. The first run must still carry the images even where the schema marks both fields optional: the member refuses a run with neither.

The anchor can be an upload: the user's character art, sketch, or product photo goes up with plus (see ) and rides the reference field like any approved hero. (see ) returns a style synthesis (medium, palette, lighting, composition, mood) that seeds the palette and lighting lines of the enumeration; the sub-element inventory comes from looking, or from with the enumeration headings as its instruction (both are catalog tools; prices with ).

A set takes one per item: a batch-count field repeats one prompt and cannot carry a per-item delta clause.

A style reference is a different slot from a subject reference, and a few image models expose it: at authoring time one family took one to ten images as , built a private style from them for a small extra charge, and returned a reusable style id in the job output that later runs pass as instead of the images (the two are mutually exclusive), with choosing or adherence. It holds rendering, palette, and line, not a character: pair it with a subject reference or the baseline enumeration for identity. Find such members with and the style need in the user's own words, then confirm the field on . Palette fields taking RGB triples () are preferences, not constraints: keep the hex values in the prompt too.

Generate the map and pass it as a conditioning input. Never extract it for the attribute that is the delta: a pose map from the approved image locks the pose you were asked to change; a pose set needs its map from a target-pose image, or none.

takes an and a from , , , , , , , , , ( defaults true). Catalog-only and write lane, despite the docs page grouping it under Analysis: run it via .

The control block (, , , , ) exists only on models listing in : check before planning around it; models without it take reference images. allows , , , , , and : only canny, depth and pose map across, becoming . defaults to 0.7 with a recommended 0.3 to 0.8 band: near 0.7 for canny, depth and tile, 0.8 to 0.9 for pose, gray and blur, rigid above 0.9. Strength is how much, and are when: near 0.65 locks composition early, then releases so the prompt refines detail.

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 , so with and the consistency need in the user's own words ("the same character as in these images"), then confirm on that the pick carries a multi-slot 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 or slot (capped between 3 and 15 seconds, billed per second of upload on some). When nothing ranked carries the array, a family name from the video model-family skills lists. Slot count moves the price less than resolution and duration: read both enums off , where 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 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. , 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 , 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 lists. Judge identity across the whole clip: returns and as free asset ids, but a drift that fades in mid-shot passes both, so sweep the frames with (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 ( is an integer count of source frames between extractions, the clip's from halved and rounded, 12 at 24 fps; the clip goes in ), 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.

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 . Dataset size, base choice, and the quote-before-launch rule: . A trained model never runs by its own id ( and , see ), 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 , where one render holds the identity instead of a generation per frame.