# Output spec ## Deliverables | File | Format | Size | Notes | | --- | --- | --- | --- | | `character.json` | JSON | | The locks, the lock text, the consent record (for a photo), the scene tests and the paths below. Read by later Skills. | | `character-sheet.jpg` | JPEG | 1792 x 1008 (16:9) | The board for people: four full-length views, four portraits, the name, the view and expression labels. Labels drawn locally. | | `character-sheet.jpg.json` | JSON | | The layout record: labels, cells, gutters, backdrop, art hashes. Read by check-sheet.mjs. | | `art/sheet-high-1.jpg` | JPEG | 1792 x 1008 | The text-free art the board was drawn on. This is `character.reference`: the file later shots attach. | | `scenes/scene-1-1.jpg` to `scene-3-1.jpg` | JPEG | 16:9, 1K or larger | Three scene tests, `nano-banana-pro` edit, the text-free sheet as the reference. | | `lock-text.txt` | Text | | The lock text, the short lock and the reference path, for pasting. | | `outfit-alt-1.jpg` | JPEG | 16:9 | Optional: the second outfit, front and back. | | `clip-1-kling-1.mp4` | MP4 | 16:9, about 5 s | Optional: the character turning to the lens, to show the face holds in motion. | | `preview.webp` | WebP | at most 2048 px, aim for 150 to 400 KB | Optional: the board for a message or a web page. | Defaults: `openai/gpt-image-2` for the sheet (`/edit` with a photo), quality "high", 1792 x 1008, `output_format` "jpeg"; a low-quality draft first. Scene tests `fal-ai/nano-banana-pro/edit`, `aspect_ratio` "16:9", `resolution` "1K". Clip `fal-ai/kling-video/v3/turbo/pro/image-to-video`, `duration` "5" (returns H.264 at about 1.4 times the still's size, with an audio track). ## Approximate cost (fal list prices, read 2026-11-13; `build-requests.mjs --stage estimate` prints the real figure) | Step | Cost | | --- | --- | | Layout draft, low quality | about $0.01 | | Sheet, high quality (about $0.01 more with a photo) | about $0.34 | | A sheet re-roll | about $0.34 | | Three scene tests, $0.15 each | $0.45 | | Second outfit (optional) | $0.15 | | Test clip, 5 s, Kling 3 Turbo Pro (optional) | $0.70 | | Seedance 2.0 mini 480p, 5 s, a face-acceptance test (optional) | about $0.36 | | Labels, lock text, checks, hand-off | free | | **Sheet and three scene tests / with one sheet re-roll / with the clip** | **about $0.80 / $1.15 / $1.50** | | The photoreal sample as run: three layout drafts, the sheet, four scene stills (one re-rolled), the clip | about $1.70 | Priced from fal's list prices and counted by the Meleon lab's ledger (its estimates per call). GPT Image 2 is billed by tokens and fal returns no token count, so the fal dashboard shows the exact charge. A call that fal refuses (the Seedance probe on a stylised face) is counted by the lab and may still be billed. ## Checklist Technical (`scripts/check-sheet.mjs`, exit code 0): 1. `character.json` passes every rule: an adult age, no named studio or artist, no real or famous person, no `FILL:`, one to six `never` items, three scenes, a lock text that matches the fields now. 2. A character from a photo has the photo path, the user's `consent_quote` and `adult_confirmed`; an invented one has none of them. 3. The board is a JPEG, 1792 x 1008, unchanged since it was composed, drawn on the art that `character.reference` names. 4. Every label (the name, the one-line who, four views, four expressions) is the exact word, inside the board, with less than 1.5% art under it and contrast of at least 4.5:1. 5. Each of the eight cells holds a figure (at least 3% of its area), and every neighbouring pair of full-length figures, and the last figure and the portraits, have a clear gap of at least 6 px between them (under 20 px is a warning). The two columns of portraits may touch. 6. Three scene tests exist and are 16:9. Optional: the second outfit exists; the clip is a video stream of about 5 s. 7. With `--preview`: a WebP, at most 2048 px, ideally 150 to 400 KB. By eye (open the guides, the board at full size and each scene test next to the sheet): 8. The same person in all eight cells: face shape, skin tone, eyes, brows, the mark. The four portraits are one person, not four. 9. The outfit exactly as locked in every full-length view, with the props, and every `never` item visible where it should be (watch on the left wrist in the front, three-quarter and back views, hair knot at the back). 10. Natural hands (five fingers, relaxed), natural feet, true-to-life proportions, no extra limbs. 11. The views are what the labels say: front faces the camera, the profile faces left, the back shows no face. 12. Nothing written anywhere: no text on clothing, a badge or the backdrop. 13. No extra people, and no backdrop shadow, floor line or panel edge in a label band. 14. Scene tests: the face, the hair and how it is worn, the outfit and every `never` item match the sheet in all three places and lights. The face is shown clearly, not only from behind. 15. The test clip, if made: the same face at the start and the end, the hair and the watch unchanged, no cut. 16. People: an invented adult resembling no real person, or the user's own photo with consent. Never a child. If a check fails: at most 3 attempts per image, change one thing per attempt and say which, compare each attempt with the sheet, then deliver the best and say plainly what still differs. Label and layout-number problems are free to fix: edit `character.json` or `layout.json` and compose again. ## Known imperfections in the sample Mara Ilves, run on 2026-11-13: one high-quality sheet (first try, no re-roll), three scene tests (scene 2 re-rolled once), one Kling clip. The technical check passes. By eye: - **Layout: the model keeps its own.** Across four boards it ignored the prompt's percentages: figures about 15% apart from x 2%, feet near 93% of the height, portraits about 17% wide and touching, whatever it was asked. `assets/layout.json` was recalibrated to those habits, and labels are set in the gap the art leaves (`references/sheet-layout.md`). In the sample the lower portraits sit very close to the board's bottom edge labels, and the labels are a few pixels out of line with each other because each sits in its own gap. - **Age and sleeves.** She reads about 40 rather than mid-30s, and the grey sleeves run to the wrist under short navy sleeves rather than being pushed to the elbows. The locks are followed in every view; a tighter age words ("early 30s") and "sleeves pushed up to the elbow" in the wardrobe would help. - **The mole** under the left eye is clear in the portraits and in the kitchen test, faint in the full-length front view and not visible in the corridor test (the face is small there). - **Scene 2 (bus stop), first try:** a hood appeared at her neck. The sample prompt said "hood down", so the model added a hooded top the wardrobe lock does not have. The prompt no longer says it; the re-roll has no hood. In the re-roll she faces right, so the watch on her left wrist is hidden. - **Scene 1 (corridor):** the face is small and a little narrower than the sheet's, but it is the same woman; the face holds in the clip. - **The clip** is a walk that settles into a steady stare, not a turn, because the corridor still already faces the lens. Pick a three-quarter scene for a real turn. No cut, the hair, outfit, clipboard and watch hold from the first frame to the last. - **Stylised 3D**, made as a $0.01 draft only to run the probe: the sheet reads well, but two pairs of figures touched (clear gaps of 2 px and 0 px), so a real stylised sheet needs a re-roll. - **Seedance 2.0 mini refused a stylised 3D start frame** ("may contain likenesses of real people"), as it did the photoreal ones. A clay face is untested.