Scenario Fan Cam

Scenario Fan Cam

Overview

A fan cam is a two-stage build: an identity-preserving image edit places the person into a 16:9 broadcast still, and image-to-video animates the approved still into a reaction. The uploaded photo is an identity reference, never a start frame: feeding it straight to a video model animates a portrait, not a broadcast. Stage one is cheap and stage two is not, so the still carries every decision worth approving.

Use only photos of the user or of people who gave them permission; refuse celebrity or stranger insertions. Real faces also trip provider filters more than most subjects: on a block, see scenario-moderation. Connection and the core loop: see the scenario skill in this repo. If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.

Quick reference

Discover members with recommend, passing the stage's capability and the user's own words: it ranks by measured cost and latency and names the purpose-built pick, where a capability-worded search returns hundreds of keyword hits with nothing to choose between them. Read next_step before taking a pick, per the scenario skill. Keep search for a member you can already name. Never assert a generative model's id as a constant. Scenario's own single-purpose tool models are named outright below: there is exactly one of each, so discovering them would only re-derive a constant.

Broadcast grammar is what sells the shot, so prompt it explicitly at both stages: long-lens compression with the crowd defocused in front and behind, harsh stadium floodlight or arena strobe, slight motion blur, the camera hunting and reframing as it finds the subject. Keep the person mid-ground among other fans, off-center, at broadcast camera height. A centered, well-lit, eye-level subject reads as a photoshoot in a stadium, not a cutaway.

Give the reaction an arc rather than a state: oblivious, then noticing the camera or the screen, then the reaction the user asked for (cheering, laughing, disbelief, heartbreak). Multi-beat prompting per the video family's contract keeps the turn inside one clip.

Worked example: caught on the stadium screen, 8 seconds

1. Collect the photo, the sport or event, venue mood, wardrobe, and the wanted reaction. upload_asset the photo once and reuse the returned asset id.
2. Create this run's collection before the first generation (collection_create, catalog write lane, name only), then collection_add_assets each keeper as it lands; its returned itemCount is the receipt.
3. recommend with capability="img2img" and the placement described in words; prefer a ranked entry that takes several reference images, and read model_schema_get for those inputs (scenario-image for the lane, scenario-consistency for reference discipline).
4. Edit prompt: "the person from the reference image seated mid-crowd at a floodlit football stadium, 16:9 television cutaway, long-lens crowd compression, no text, logos, or graphics in frame". Keep the plate text-free: overlays come later.
5. Gate the still: asset_display it for approval, and inventory the likeness features that must hold (face geometry, hairline, skin tone, wardrobe) with the analysis lane from scenario-asset-analysis; a change in any inventoried feature fails. Unattended, that inventory stands in for the user's sign-off.
6. recommend with capability="img2video" and the reaction described in words, then check the pick against scenario-kling or scenario-video. The brief's length filters members: duration is usually a fixed enum a member either reaches or does not, and the tiers that reach it sit far apart on price, so model_schema_get each candidate and compare with model_run dry_run=true.
7. Run with the approved still as the start image and two beats: "she chats, unaware" then "she spots herself on the stadium screen, stands, and cheers", one slow reframing move, wait=false, then jobs_wait re-called with pending_job_ids on timeout, never a second model_run.
8. Check the clip against the approved still on frames the platform produced: the free firstFrame and lastFrame off asset_get first, then model_scenario-video-to-image-seq for a mid-clip frame (a tool-model lane per scenario-video-editing). A local extraction is not a substitute at any budget, because it yields no asset to file or audit and the gate stops being traceable. A change in any inventoried feature fails the clip, and so does legible generated type: ribbon boards and jumbotron glyphs creep in even against a no-text prompt, worst late in the clip. The retry starts from the same still, never from the drifted output.
9. Assemble on-platform with model_scenario-compose-video (a tool-model lane per scenario-video-assembly), layering the score strip and channel bug from scenario-text-overlay in the safe corners; a local editor is the wrong lane. Give every overlay layer an explicit width and height, since leaving them empty is documented as native size and does not hold, and pin durationMode: "custom" to the clip's length, since an image layer otherwise stretches the composition past it. A succeeded compose job is not proof the layers drew: pull a frame from the composite (the free firstFrame) and confirm every layer is present and placed before delivering. Then cut 9:16 per scenario-formats, a resize rather than a reframe, whose video members outpaint a wider frame instead of cutting one. Its cover mode crops from the center, so when the subject sits off-center put the master on a 9:16 compositor canvas positioned to hold them instead. asset_display both.