Can AI generalize a filmmaker's visual grammar to scenes they never actually shot?

Yes — when a director's visual grammar is codified as a structured system (camera, lens, palette, lighting, composition, mood, negative rules) and loaded as persistent context, an AI agent can apply that grammar to scenes the director never shot. The proof is whether the agent makes the right choice on a genre or moment outside the source body of work.

Generalization only happens when the style is encoded as a decision system, not pasted as reference frames. The invideo agent is an agentic video tool that holds a treatment document in persistent context and routes generation across the current model roster (Runway, Veo, Kling, Seedance 2.0 for video; Recraft, Nano Banana, GPT-Image-2 for images), so the grammar travels with every shot instead of getting re-prompted scene by scene.

Encode the grammar as rules, not vibes. A documented Wong Kar-wai build ran 14 sections — camera, angles, colour tone, atmosphere, mood, lighting, composition, movement, film palettes, prompt templates, negative prompts, quick-reference card — into a 25-page treatment. A James Wan build added a five-stage emotional architecture (each stage with locked rules for camera, lighting, sound) plus an 85:15 dark-to-light ratio and an explicit exceptions directive that quarantines outlier films so the agent doesn't misapply general rules. Rules with named modes (e.g. tonal modes with exact hex values) generalize; mood-board screenshots do not.

Validate by testing the grammar on something the director never made. The reliable test is to ask the agent for a scene type outside the director's filmography — a courtroom thriller through the Wan lens, a non-horror moment in his style — before generating a single production frame. If the agent asks clarifying questions (era, nature of the threat) and applies named principles from the doc to the new scene type, the grammar is internalized. In one documented case the agent pulled "Mood Over Narrative — the substitution rule" from page 12 of the treatment and applied it to a scene the document had never specifically addressed, then independently recommended a "doorway static hold" ending — a structurally recurring device across two Wong Kar-wai films the agent connected on its own.

Watch for grammar-level behaviour during production, not just style-level output. Real generalization shows up as the agent reasoning ahead and self-correcting against the doc. Documented examples: the agent flagged that Stage A shadows were leaning blue-green instead of neutral gray and offered a warmer pass without being asked; caught that an entity-reveal shot was running at the wrong emotional stage register (Stage D instead of C) the human director had missed; corrected its own cinematography note from anamorphic to spherical when challenged on lens type; built a precise reverse angle with no reference image, using only geography established in earlier shots. Those are decisions made from grammar, not pixels matched from references.

Where it still falls short. Generalization tracks visual grammar reliably; it tracks narrative pacing, emotional rhythm, and sound design only as well as those are explicitly written into the doc. The Wan build worked partly because the treatment included a full audio architecture module — "half of what makes his films land is what you hear before what you see." If pacing and sound logic aren't encoded, the agent will hold camera and palette across new scenes but drift on rhythm. Cross-genre transfer also needs an exceptions section so the agent doesn't flatten everything to the dominant mode.

The practical handoff. Hand the agent: visual grammar (camera, lens, palette, lighting, composition, movement), named modes with exact values, emotional-stage rules with "what never to do" per stage, negative prompts, and any audio logic that shapes shot choice. Keep for yourself: final emotional-register call on key beats, pacing decisions on a rough cut (then send the cut back as a maker-checker pass against the doc), and the call on whether an "off" suggestion is wrong or right. Across documented productions in this style — a 70-second Wong Kar-wai short, a ~90-second Wan-style horror short, and a three-film director series — the consistent pattern is that the agent generalizes faithfully on whatever was written down with rules, and improvises weakly on whatever was left implicit.

Watch some of these to see what works for you:

that's the moment I knew the doc was solid because the agent understood James Wan's grammar well enough to apply it to even a non-horror moment.

— Hridaye, invideo's creative director