Why does prompting each scene separately cause style drift in AI filmmaking?

Per-scene prompting drifts because video models have no memory between generations — every new prompt is a cold start, so style is re-sampled instead of retrieved. Small wording changes resample lighting, palette, lens, and texture. Style is a film-level constant, not a scene-level variable, and prompting it scene by scene treats it as the latter.

The root cause is statelessness. A generation model doesn't remember what it made five minutes ago — each prompt is rendered from scratch against the model's prior, so any directive you don't repeat verbatim gets resampled. Rewrite "warm amber light" as "warm yellow lamps" in scene 4 and the model legitimately reads that as a new instruction; lighting shifts, palette shifts, skin texture shifts. Multiply that across 20-160 prompts in a film and the drift compounds into a visibly different look by act three.

The creative error sits underneath the technical one: per-scene prompting treats style as something you describe again every scene, when style is the one thing that must NOT change. Camera grammar, palette, lens character, atmosphere, mood register — these are film-level constants. Scene-level variables are blocking, action, and emotional beat. Mixing those two layers into one prompt per scene is the anti-pattern; every prompt becomes a fresh negotiation with the model over what the film looks like.

Prompt wording isn't the only drift vector. Model choice drifts (Runway vs. Kling vs. Seedance 2.0 each carry different priors for lighting and skin), seed drifts, and reference-image selection drifts — if you swap which still you attach scene to scene, the model re-anchors to whatever's strongest in the new image. As one creative director put it: "Agent One reads your treatment doc once and keeps it loaded across every frame. The thread stays held, scene to scene. No re-explaining. No starting over." — Hridaye, invideo's creative director. That's the architectural fix: lift style out of the per-scene prompt and into a persistent context the model sees every time.

The practical fix is to externalize the style as a locked block — camera spec, lens, lighting source, palette with hex values, composition, atmosphere, mood register, film/DP attribution, negative prompt — written once and prepended unchanged to every generation. Across documented productions a fixed 9-element prompt assembly order is held across every frame, with 14 codified principles (camera, angles, colour tone, atmosphere, mood, lighting, composition, movement, palettes, templates, negatives, quick-reference) stored as a 25-page treatment loaded once. In one production the discipline was that "every prompt after this started with it" — the style block appeared on 100% of generations, including the 64 reference frames first ingested with "deeply understand this art style and save it into context for further generations." Negative constraints matter as much as positive ones; explicit prohibitions like "not live-action, not photorealistic" prevent the model from drifting toward its default prior between scenes.

invideo is an agentic video creation platform with every current video and image model and the upscalers available behind one agent. Practically, this means you load the style document into the invideo agent once and it routes each shot — to Veo, Kling, or Seedance 2.0 depending on what the shot needs — while holding the style block constant across the route. The agent becomes the persistence layer the underlying models lack: same style directives, same negative prompts, same palette references attached to every generation, whether shot 3 goes to one model and shot 47 goes to another. You vary only what should vary per scene — blocking, beat, emotion — and the film-level constants stay constant by construction.

Watch some of these to see what works for you:

Agent One reads your treatment doc once and keeps it loaded across every frame. The thread stays held, scene to scene. No re-explaining. No starting over.

— Hridaye, invideo's creative director