AI Character¶

One of six LoRA types on Lookhouse. The one that gives a performance a person.

In one line¶

An AI Character is a small, trainable add-on that teaches an AI diffusion model to recreate one specific person in motion — the same face, the same body, and the same way they move, react, and emote — on demand, in any scene you can describe.

Without it, AI gives you a different stranger in every clip. With it, you get the same character, performing — frame after frame, cut after cut.

Why video is the hard part¶

A still image only has to get a face right once. Video has to get it right twenty-four times a second — and keep it right while the person moves.

That's a different problem. A human isn't just a set of facial features. A human is a set of behaviors: the half-smile that starts on one side of the mouth, the way the eyebrows lift a beat before a laugh, the tilt of the head while listening, the rhythm of a blink, the small gestures that fill a pause. We recognize people we know from across a street by the way they walk — long before we can see their face.

This is what makes character consistency in video so brutal. Ask a raw AI diffusion model for "the same person" across two clips and you don't just get a wandering jawline — you get a different human being, because the expressions and movement don't carry over. The face might pass; the performance gives it away. The result feels like a stranger doing an impression of your character.

An AI Character fixes exactly this. It learns the person as a moving, emoting whole — not a photo, but a presence.

The problem it solves¶

Base AI diffusion models (the engines behind AI video) are brilliant at "a woman with red hair, smiling" and useless at "her — the one from the last clip, smiling the way she smiles."

Ask a raw model for the same person twice and you get two different people who happen to share a hair color. Different micro-expressions. A new way of holding the eyes. Gestures that belong to no one. For a single generated clip, you might not notice. For a film, a series, a campaign that needs one continuous performance — the audience notices instantly, even if they can't say why.

An AI Character locks the whole identity in place — face, body, and behavior. Train it once on a person, and that person becomes a reusable performer you can direct into any scene: different lighting, different wardrobe, different city, different decade — same recognizable human, moving and reacting like themselves.

This is the difference between generating clips and casting talent.

How it works¶

You don't need the math to use it, but here's the honest version of what's happening under the hood.

The base model already knows almost everything. It understands faces, fabric, sunlight, motion, depth. What it doesn't know is your particular person. Retraining the whole model to learn one face would cost a fortune and take forever — these models have billions of parameters.

LoRA (Low-Rank Adaptation) is the shortcut. Instead of rewriting the entire model, LoRA freezes it and trains a tiny set of new weights that "nudge" the model toward one specific identity. Think of it as a thin overlay clipped onto a finished engine — the engine does the heavy lifting, the overlay steers it toward one face.

Because it only learns the difference between "a generic person" and "this person," an AI Character is small (often just a few megabytes to a few hundred), fast to train, and easy to share or sell as a single file.

Why video input matters. Images can only teach a model what a person looks like. Video teaches it what a person is like — how the face moves between expressions, the timing of a smile, the way the head and hands behave during speech and reaction. When you train on video, the AI Character captures not a frozen likeness but a living one: the behavioral signature that makes a person recognizable in motion. This is why Lookhouse builds AI Characters from video, not just stills.

The training loop, in plain terms:

You feed it footage of one person — ideally video, so the model sees them move and emote, plus supporting stills.
The system studies what stays constant across every frame — the bone structure, the proportions, the signature features, and the recurring patterns of expression and movement — and separates it from what changes (background, pose, outfit, lighting, the specific words being said).
It writes that "constant identity" — looks and behavior — into the LoRA weights.
From then on, whenever the LoRA is active, the model bends its output toward that identity, keeping the person consistent from frame to frame while still obeying everything else in your prompt.

A few terms you'll see on Lookhouse:

Base / checkpoint — the foundation model the LoRA is built to run on (e.g. an SDXL- or Flux-family model). A LoRA is matched to a base; using the right one matters for quality.
Trigger word — a unique token (like a stage name) you put in the prompt to call the character into the scene.
Rank — roughly, how much "room" the LoRA has to store detail. Higher rank can capture more nuance but makes a larger file; well-tuned lower ranks are often sharper and more flexible.
Strength / weight — a dial at generation time. Turn it up to push harder toward the exact likeness; ease it down to let the character blend more naturally into a new style.

The part most people miss: expression and interaction style¶

Two people can have nearly identical faces and still be unmistakable as individuals. Why? Because identity lives in motion as much as in features.

Every person carries a behavioral fingerprint:

Expression signature — how they smile, frown, react, and emote. The asymmetry of a grin. The squint before a laugh. The face they make when they're thinking. Most people have a handful of expressions they return to constantly, and that repertoire is as unique as a voice.
Micro-movement — blink rate, the way the gaze drifts and settles, the small involuntary tics that read as "alive."
Interaction style — how they hold themselves while listening, the gestures that punctuate their speech, how they lean in, look away, nod, fidget. The body language of being this person in a room with someone.
Rhythm and timing — the pace of their reactions, the beats between a stimulus and a response. Two people can make the same gesture; the timing of it is what we recognize.

A face-only model — or an AI Character trained only on stills — can copy the features and miss all of this. The output looks like the person in a freeze-frame and feels like an impostor the moment they move. It lands in the uncanny valley not because the face is wrong, but because the behavior is generic.

An AI Character trained on video learns this layer. It doesn't just reproduce a face; it reproduces a performer — someone who reacts and emotes the way the real person does. That's the difference between an audience accepting your character and an audience squinting at the screen, sensing that something is off.

For a creator, this is the real value of your likeness: not just your face, but your presence — the thing casting directors actually hire. For a producer, it's the difference between a clip that holds up in a finished edit and one that breaks the illusion.

What you get from it¶

Consistency across every frame. The single biggest reason AI Characters exist. Same person across a full sequence of clips — same look and same way of moving, reacting, and emoting. No drift, no morphing, no "wait, is that the same actor?"

Infinite reusability. Train once, use forever. The character can perform in a noir alley, a sunlit kitchen, a sci-fi cockpit — costume changes, age shifts, lighting moods, new lines of action — all from text prompts, all recognizably the same person, reacting like themselves.

Speed and cost. Minutes-to-hours to train instead of the days and budget a full fine-tune demands. A few megabytes instead of gigabytes. Cheap to store, instant to load, trivial to share.

Control. The strength dial, trigger word, and prompt give you director-level command — dial the likeness up for a close-up, down for a stylized wide shot.

Portability. It's one file. It moves. That's what makes it sellable — and what makes Lookhouse a marketplace, not just a tool.

Who this is for on Lookhouse¶

If you're a Creator — an actor, actress, or model¶

Your face is your asset. An AI Character turns it into one you can license at scale.

You upload your video and images; Lookhouse builds an AI Character trained on you. From that point, your likeness and your way of moving can be cast into projects you'd never have had time or budget to shoot in person — and you set the terms. You're not selling a photo. You're selling a digital version of you that performs, while the original stays yours.

What to bring for a strong model: video first — clips that show you talking, reacting, and emoting naturally, so the model learns your expressions and timing, not just your features. Add a range of clear stills across varied angles, expressions, and lighting, with your features unobstructed. Variety in behavior is what captures the real you; variety in lighting is what stops the model confusing you with the room you were standing in. (Lookhouse handles the training; your job is good source material.)

If you're an AI Developer¶

You already know the training game. Lookhouse is where your work gets paid.

Upload the AI Characters you've trained, set your price, and reach buyers who need exactly this — identity-consistent models for production, not experiments. List your base model, rank, recommended strength, and trigger word so buyers can deploy in seconds. Good documentation sells; an AI Character that "just works" on the stated base earns repeat buyers.

If you're a Buyer — an AI film producer¶

You need the same character to hold up across a whole project. That's the entire pitch.

Download an AI Character — built by a creator from their own footage, or trained and listed by a developer — and drop it into your pipeline. Now your lead is consistent across every shot, scene, and clip. No re-shoots, no continuity errors, no negotiating a second studio day. Cast from a library instead of a calendar.

AI Character vs. the other LoRA types¶

AI Character is one of six LoRA types on Lookhouse, and it answers a different question than the others:

An AI Character answers "who is in the shot?" — it locks an identity.
The other types shape the how and what of the shot — the look, the medium, the wardrobe, the setting, the action.

The power move is stacking: load an AI Character for who, and combine it with other LoRA types for everything else. Same person, infinite worlds. That combination — a consistent identity dropped into any style or scene — is what turns a pile of models into an actual production toolkit.

The honest limits¶

Good documentation tells you where the edges are.

Quality follows the source. Blurry, repetitive, or single-angle input gives a rigid, fragile model. If the footage only ever shows one expression, that's the only expression you'll reliably get back. Range of behavior in, range of behavior out.
Likeness vs. flexibility is a trade-off. Push the strength too high and the character becomes stiff and hard to restyle; too low and the resemblance slips. The sweet spot is part of the craft.
Base-model match matters. An AI Character trained for one base family won't behave correctly on another. Always check the listed base before buying or deploying.
Consent and rights are non-negotiable. An AI Character is a person's likeness. On Lookhouse, that means clear ownership and permission — creators license their own face; buyers get models with rights attached. This isn't fine print; it's the foundation of a marketplace built on real people.

The takeaway¶

An AI Character is the smallest file that solves AI video's biggest problem: keeping a person a person in motion. It captures not just a face but a performance — the expressions, the timing, the way of reacting that make someone recognizable — and turns it into a castable, sellable, reusable asset. Train once, perform anywhere.

For creators, it's your presence, working while you sleep. For developers, it's your craft, finally priced. For producers, it's a casting library that never misses a continuity note — or a beat.

One person. Every frame. Performing like themselves.