How can you tell if an AI has truly learned a director's visual style or is just copying it?

Test for abstraction, not replication. A genuinely trained agent applies the director's grammar — lighting ratios, lens behavior, palette logic, blocking — to scenes the director never shot, asks clarifying questions before generating, and self-corrects technical claims when challenged. A copying agent reproduces referenced frames, collapses on unfamiliar briefs, and never questions you back.

Run a three-part diagnostic on the agent holding your director's visual language document.

1. The novel-brief test (genre transplant). Ask the agent to apply the director's style to a genre or subject that director never worked in — a courtroom thriller through James Wan's lens, Wong Kar-wai applied to a sci-fi interior. If the agent has truly internalized the grammar, it returns shots that are stylistically coherent in lighting ratio, lens choice, palette mode, and blocking — without you re-stating any of those. If it has only pattern-matched, it either refuses, returns generic genre output, or leans on referenced frames that don't fit. One documented horror production stress-tested its director bible this exact way before generating a single frame: "Before generating a single frame, I stress-tested the doc. I asked for a courtroom thriller through the James Wan lens. Something he's never made. If the agent was just mirroring style superficially, it would fail here."

2. The clarifying-question test. A trained agent surfaces ambiguity before it builds the frame — it asks you what era, what entity reference, what prop, what deliverable format, what's behind the character in the reverse. A copying agent assumes and hallucinates a resolution. Watch what it does when you give it an underspecified brief: questions back = grammar internalized; silent generation that defaults to the nearest reference = mimicry. The same horror production logged the agent's pre-generation question set — character, entity, toy, deliverable — before any asset was made; that's the behavior signature you want.

3. The challenge-and-correct test. Question the agent's technical claims about the director's craft — lens type, aspect ratio, lighting source, dark-to-light ratio. An agent holding the language as grammar will check the document and self-correct (one production caught the agent calling The Conjuring anamorphic and watched it correct to spherical, 35mm, 2.40:1 hard matte when challenged). An agent that's mimicking will agree with whatever you push, because it has no internal rule set to defend.

Two supporting signals that strengthen the diagnosis. Proactive rule application: the agent flags deviations you didn't ask it to check — "I was generating Scene 1 and before I noticed anything, the agent caught that the shadows were leaning blue-green instead of neutral gray. Pulled the Stage A rule from the doc, flagged the deviation, offered a warmer pass. I never asked it to cross-check." That's grammar in action. Structural recall on novel asks: when you ask for something the document never explicitly addressed (an ending you can't write, a reverse angle with no reference), a trained agent pulls a named principle from a specific page and applies it — one production's agent surfaced the "substitution rule" from page 12 and a recurring "doorway static hold" device, sequencing six closing shots autonomously. Pattern-matchers cannot do this; they need a referenced frame in front of them.

A quick mimicry check on the image side: reverse-image-search any character sheet or hero frame the agent produces. If it pulls back near-duplicates of the source references rather than novel compositions that share the grammar, you have a copy problem, not an abstraction.

One note on tooling context: the invideo agent reads your treatment document once and holds it across every shot, which is what makes these tests meaningful — you're evaluating whether the loaded grammar generalizes, not whether a single prompt landed. Hridaye, invideo's creative director, frames the bar this way: "AN AGENT IS ONLY AS POWERFUL AS WHAT YOU TEACH IT. We taught one Wong Kar-wai. You decide who your agent learns from next." The diagnostics above tell you whether the teaching took.

If the agent passes the novel-brief test, asks before it builds, and corrects itself when challenged, it has internalized the visual language as grammar. If it copies referenced frames, agrees with everything, and collapses on unfamiliar briefs, you have surface mimicry — go back and rebuild the document with explicit rules per stage, per shot type, and per emotional register before generating further.

Watch some of these to see what works for you:

Before generating a single frame, I stress-tested the doc. I asked for a courtroom thriller through the James Wan lens. Something he's never made. If the agent was just mirroring style superficially, it would fail here.

— Hridaye, invideo's creative director