Stop Generating Posters

Types vs. People

There is a default the model falls into when you describe a character in broad terms. The heroic weathered face. The perfectly groomed beard. The direct cinematic gaze. These aren't characters - they're casting archetypes. They look like someone a studio would hire. They don't look like someone you'd pass on the street.

The camera was pointed at them. They weren't found.

This is the distinction that matters. A found person exists before the camera arrives. A poster subject exists because the camera is there. AI generation defaults to the second mode. The prompt has to actively fight it.

The fix is not to describe characters less. It's to describe them differently - with the kind of specific, slightly unflattering detail that only exists in reality.

What to Change

Look off-camera. This is the single most effective change you can make. Direct gaze triggers hero mode - the model reads it as a portrait and gives you a portrait subject. An off-camera gaze - eyes focused on something just out of frame, attention somewhere else - makes someone a person in a room rather than a subject in a photograph.

Undercut the beard. The moment you describe facial hair with precision - dense, heavy, perfectly maintained - the model generates something groomed. "Stubble, four or five days, patchy and uneven at the jaw" reads as real. Specific imperfection defeats generic grooming.

Replace mood adjectives with behavioral specifics. "Intense, weary eyes" unlocks a stock archetype. "Tired that's become the default" produces the same exhaustion without triggering the archetype that owns the word. The behavioral description is harder to romanticize.

Give the face something small to do. "Lips parted slightly, breath just barely visible." A face with nothing to do defaults to neutral heroism. A face doing something specific - even something minimal - becomes a person caught in a moment rather than posed for one.

Undercut the wardrobe. Clothing that fits perfectly reads as costume. Add history: "slightly misshapen from years of use." The jacket that's been worn enough to lose its shape belongs to someone. The jacket that fits reads as selected.

Delete intensity descriptors entirely. Remove "intense," "gritty," "weathered," "powerful," "commanding" from the prompt. What remains when those words are gone is quieter and more real. The intensity descriptors are the words that produce posters.

Same character. The second description finds a person. The first generates a poster.

Test Before You Build

Before generating a full character reference sheet, run a single portrait close-up with an off-camera gaze. One generation. Then look at it honestly: does it look like someone you might pass on the street, or does it look like someone cast in a film?

If the answer is cast - fix the description before building the sheet. This is not a small thing. A reference sheet generated from a poster portrait produces poster consistency across every shot that uses it. The error compounds from the first generation forward. Every subsequent shot inherits the archetype, locked in at the level of the reference.

The test costs one generation. Fixing it after the sheet is built costs everything downstream.

Why This Matters for Narrative

A poster subject can carry a genre film. They look the part. They signal the right things. For AI filmmaking that aims at something more specific - a character with a particular interior life, a face that carries history rather than archetype - the poster is the enemy of the work.

The characters in The Last Analog had to be specific people, not types. A scientist's daughter returning to finish work her father left incomplete is not a hero archetype. She has a particular relationship with exhaustion and with cold and with the kind of focused attention that comes from years of training. None of that lives in "intense, weary eyes." All of it lives in the behavioral specifics.