--- name: character-and-continuity version: "0.1.0" description: The casting director's, the costume designer's and the script supervisor's craft for narrative films - the locked look sheet per character (one front three-quarter portrait generated once, every frame edited from it), wardrobe per scene, the continuity record (props, hair, wardrobe, screen direction and the 180-degree line, eyelines, light), and the identity checks the quality gate applies. Use when a series bible's characters are cast, before any frame with a person in it is generated, when a scene changes wardrobe or props, when a frame is regenerated, and before the edit of every episode. license: BUSL-1.1 metadata: author: 贺家豪 (Jiahao He) category: production tags: - casting - costume - continuity - micro-drama - qwen-image-edit --- # Character and continuity In an ad the product hero still is generated once and every frame is edited from it, so the product never drifts. A narrative film does the same with people: each character has one locked portrait, and every frame that shows them is an edit of that portrait. Viewers forgive flat light; they do not forgive a lead whose jaw, hair parting or jacket changes between a shot and its reverse. Audience complaints about AI dramas rank "stiff faces" and "odd movement" first (research, 3.4); a face that changes is worse than either. ## Casting rules - Adults only: every character is 18-90 (*checked*). Write the age in words in `look` ("in her late twenties"), never digits (*checked*); the image model paints numbers. - No real people. Never name or describe a celebrity, a public figure, a client's employee or "someone who looks like" anyone. Unauthorised likeness is the industry's main legal exposure (research, 3.4; NO FAKES Act passed committee 2026-06). - At most five characters by default (*checked*): each needs a locked look and a designed voice, and a 10-20 episode season cannot develop more. - `look`, `wardrobe` and `voice` are English (*checked*); `id` is a lowercase slug (*checked*) used in file names, shot lists and voice files. - Cast for contrast. Two leads must be told apart in a thumbnail at phone size: different hair length and colour value, different silhouette, different wardrobe palette. Two dark-haired men in dark suits read as one person on a 6-inch screen. - Write `look` as a casting director describes a face, specific and plain: face shape, brow, eyes, nose, mouth, skin tone and texture, hair (length, parting, texture, colour), build, one distinguishing detail (a small scar through the left eyebrow, a mole below the right eye). No adjectives of worth ("beautiful", "handsome"): they pull every face toward the same model average. ## The look sheet One sheet per character, generated in pre-production and never regenerated mid-season. 1. **Master portrait.** Qwen-Image, no reference: head and shoulders, front three-quarter (face turned about thirty degrees off camera, eyes toward the lens side), neutral mid-grey seamless backdrop, large soft key from camera left with gentle fill, neutral expression, mouth closed, the default wardrobe visible at the collar. Words only, no numbers, no lettering. Two to four seeds; the director picks one; it is saved as `refs/characters/.png` and is the identity of record. 2. **Derived views, all edited from the master** with Qwen-Image-Edit ("Image 1 is : keep the face, hair and skin exactly as in image 1"): profile left, full-length in default wardrobe, and one per scene wardrobe change. Never generate a derived view from text alone - it is a different person. 3. **Gate the sheet.** Each derived view passes `shot-quality-gate` against the master before it is used: same face shape, eye spacing, hairline and parting, skin tone, distinguishing detail present on the correct side. 4. **Lock.** The sheet is read-only for the season. A recast (rare, director's call) regenerates the master and every frame of that character that is not yet in a delivered episode. ## Every frame with a person is an edit - First frames of shots that show a character come from Qwen-Image-Edit with the master portrait as image 1. At most three references (product or character first, as in `spark-local-media-generation`): the speaker's portrait, then a second character's portrait for an over-the-shoulder, then the previous approved frame of the scene for set and light. - Say what each reference is, every time: "Image 1 is Mia: keep her face, hair and skin exactly as in image 1. Image 2 is the café as established: keep the window on the left and the warm morning light." - Never re-describe the face in `image_prompt`. Describe pose, framing, expression, action and wardrobe change only; re-describing features drifts the face exactly as re-describing a product drifts the label. - Expression is directed in the prompt in small, physical terms - "jaw set, eyes wet but holding" - not with emotion labels ("sad", "angry"), which produce stock theatrical faces. - Across episodes, re-anchor every episode from the master portrait, never from the last frame of the previous episode: chaining edits compounds drift. Long seasons drift anyway (research, 5.3); keep first seasons at 10-20 episodes. ## Wardrobe per scene - `Character.wardrobe` is the default. A scene that changes it records the change in the continuity record with the scene number, and the derived wardrobe view is made once and used for every shot of that scene. - Wardrobe is specific enough to repeat: garment, colour, fabric, fit, what is open or rolled. "Charcoal wool overcoat, open, over a cream roll-neck" repeats; "smart casual" does not. - No logos, no printed text, no brand marks on clothes. Jewellery and glasses are continuity items: if she wears a thin gold ring in scene 2, she wears it in every shot of scene 2. - Story time drives changes: the same day keeps the same clothes. A change needs a reason the audience can infer (next morning, a gala). ## The continuity record The script supervisor's notes, kept per episode at `projects//continuity/ep.json`, written before frames are generated and updated after every gated frame. ```json { "episode": 3, "scenes": [{ "scene": 2, "heading": "INT. CAFÉ - MORNING", "time": "morning, overcast", "key_light": "soft window light from frame left", "line_of_action": "mia frame left facing right, leo frame right facing left", "characters": [{ "id": "mia", "wardrobe": "charcoal wool overcoat open over cream roll-neck", "hair": "shoulder length, side parting on her left, tucked behind right ear", "props": ["paper coffee cup in right hand"], "state": "eyes slightly red from scene one" }], "props_on_set": ["two cups on the table, laptop closed"], "shots": [{"shot": 4, "frame": "frames/shot_004.png", "eyeline": "to frame right, slightly above lens", "notes": ""}] }] } ``` What it tracks, because these are what break first: hair parting and tucks, which hand holds which prop and how full the cup is, jewellery, wardrobe state (buttoned, sleeves rolled), injuries and make-up states that carry across scenes, set dressing that appears in more than one shot, and time of day with key direction. ## Screen direction and eyelines - **The 180-degree line.** In a two-person scene, decide who is frame left and who is frame right in the wide, record it as `line_of_action`, and keep it: singles and over-the-shoulders show each character looking the same way across the cut. Crossing is allowed only on screen (a character walks across, the camera dollies across) or through a neutral shot straight down the line. - **Eyelines.** A single of A looking frame right, slightly above lens, cuts with a single of B looking frame left at the matching height. The taller character looks down, the shorter up, consistently. Record the eyeline per shot and write it into `image_prompt` ("looking off to frame right, just past the camera"). - Lip-sync constrains framing: a dialogue shot frames its speaker first, in medium, medium close-up, close-up or over-the-shoulder, and only the speaker faces camera (*checked*). Plan the coverage of every exchange as single / reverse single or over-the-shoulder pairs, never a two-shot with both faces talking (see `dialogue-and-voice`). - **Light direction** follows the room, not the camera: if the window is frame left in the wide, it lights A's right cheek in her single and B's left cheek in his reverse. ## Identity checks for the gate The vision reviewer (`src/crafts/narrative/gate.ts`) judges two things; a frame passes only with `pass`, score ≥ 80 and light, composition and realism each at or above the craft floor (6 by default). - **Sheet gate** on the master portrait rejects (*checked* by the judge): not exactly one person, not clearly an adult of the written age and gender, a face or hair that contradicts `look`, distortion or malformed hands, a CGI or beauty-filtered look (waxy, poreless skin, glossy eyes, perfect symmetry), any text or logo, a face that is cropped or not clearly lit. - **Shot gate** on every first frame, with up to two character portraits as references, rejects (*checked* by the judge): a character who is a different person from their portrait (another face, hair colour or style, age or build), more people facing camera than the shot lists or an unexpected extra person, the speaker's mouth hidden or soft on a lip-synced shot, distorted faces or hands, any lettering, a CGI look, flat light or a tilted, cluttered frame. The shot gate deliberately ignores wardrobe changes the scene calls for, so it cannot tell a correct change from a continuity error. Those checks are yours, against the continuity record, on every gated frame: - Hair parting, length, tucks and colour as recorded for the scene; the distinguishing detail on the correct side. - Wardrobe garment, colour and state as recorded; no logo or lettering on clothing. - Props in the recorded hand, at the recorded state; no unrecorded prop. - Screen direction and eyeline consistent with `line_of_action`. - Background faces soft or absent (crowds degrade - research, 5.3); hand contact scripted around where possible. Two candidates per frame (more with `candidates`); keep the passing one with the highest score. After a clip is generated, check its last frame too: video models drift the face over five seconds, and a shot whose face changes by the end is regenerated even though its first frame passed. ## Common failures - A derived view generated from text instead of edited from the master: a cousin, not the character. - Chaining edits frame to frame across a scene or an episode: drift accumulates; always edit from the master. - Re-describing the face in each shot prompt: features average toward a generic face. - Two similar-looking leads: the audience loses who is who in the first episode. - Continuity record written after the frames: the frames become the record, and the errors with them. - Crossing the line between a single and its reverse: both characters appear to face the same way.