LordVayron/video-factory

Cinematographer

Read references/cinematography.md (film grammar, look) and references/models.md (models) once per session.

Think in scenes, not shots

Research (Higgsfield Popcorn, Wan 3.0 / Seedance 2.0 multi-shot): AI models keep continuity when a scene is generated together, and lose it when each shot is generated alone. So:

• Scene = one location, one continuous moment, ≤ 15 s, 2–5 shots of 2–5 s each.
• Storyboard = ONE grid image per scene (all panels in one generation) → split into panels.
• Video = ONE multi-shot generation per scene (Shot 1 [0-3s] … Hard cut. Shot 2 [3-6s] …), opening on the scene's first panel, with character and set references and voice references.
• A 20 s episode is typically 2 scenes → 2 grid images + 2 video generations, all run in parallel.

shotlist.json

"scenes": [
  { "id": "A", "summary": "one line: what happens", "shots": ["s01","s02","s03"],
    "blocking": "Milo sits on the LEFT chair at the round table facing right, paper in front of him; Rex sits on the RIGHT chair facing left; Ana is not in the room." }
],
"shots": [ { "id": "s01", "duration": 3, "characters": ["milo","rex"], "framing": "medium two-shot", "lens": "35mm lens", "camera": "fixed camera",
             "image": "frozen moment for the storyboard panel", "motion": "what changes in this shot",
             "dialogue": [{ "char": "milo", "text": "…", "delivery": "…" }], "sfx": "…" } ]

Blocking sheet (write it before the shots)

For every scene: each character's position (which chair, left/right of frame), facing, posture, props in hand, and any movement with its timing ("in s03 Rex stands up and moves to the doorway"). If a character moves, the next shot starts where they ended. The blocking text goes into the storyboard grid prompt and the video prompt verbatim.

Coverage that keeps continuity

• Anchor with the two-shot. Dialogue between two seated characters plays mostly in a medium two-shot; cut in for reactions and punchlines.
• Singles show the geography: a close shot of one character keeps the table and the other's shoulder or edge in frame (over-the-shoulder), against the same background wall every time.
• 180° rule: screen sides never flip. Write "Milo frame left, Rex frame right" in every shot that has both.
• Name the camera in every shot ("fixed camera", "slow push-in"). Unnamed cameras make Wan cut on its own.
• Physical comedy in the wide, so the whole action reads.
• Cut on action or on the punchline word; keep the reaction beat (0.5–1 s) after the line.
• No impossible jumps: a character can't be at the table in s02 and in the middle of the room in s03 unless a shot shows them moving.

Before handing off

• Every scene ≤ 15 s with 2–5 shots; the total matches the script runtime.
• Every shot has framing, lens, camera, image, motion, dialogue (or none) and sfx.
• node engine/gen.mjs estimate <slug>.

Continuity logic (hard rules, enforced by engine/continuity.mjs)

• Every scene declares seats. Each character gets a place and a screen side: { "rex": { "place": "the RIGHT cushion of the couch", "side": "right" } }. Prompts repeat it as a SEAT LOCK in every shot, wides included.
• Nobody changes place off screen.
  • If the story needs a character somewhere else, add a shot whose moves says where to and how, and whose action shows the movement: "moves": { "rex": { "to": "the rug", "side": "right", "how": "hops down off the couch" } }.
  • The plan check fails if a later scene starts a character somewhere new without that shot.
• Singles are dirty singles. When characters share furniture, a close-up keeps the partner's edge (fur, ear, shoulder) at their screen side. The seat beside the subject is never empty.
• After filming, run node engine/continuity.mjs <slug>.
  • A vision judge reads one frame per shot and compares it with the plan: furniture/floor, side, teleports, night or day, empty partner seat, room, wardrobe.
  • A failing take can't be cut (edit.mjs refuses it). Reroll it with a note naming the violation.