AI Location¶

One of the six AI model types on Lookhouse. Build a place once. Shoot in it forever.

What it is¶

An AI Location is a trained model of a single real place — an apartment, a café, a nightclub, a rooftop, a back alley — that lets AI film tools generate that exact space on demand, in any light, any weather, any crowd.

Scout a location once. Capture it properly. From then on, the place is yours to shoot in at 3 AM or high noon, empty or packed, sunny or storming — without ever going back, renting it again, or matching the lighting by hand.

Technically, it's a LoRA: a small (~200 MB) model file that teaches a base video model one specific thing — here, the identity of one location — without retraining the whole model. On Lookhouse, we call it an AI Location because that's what it does for you: it gives you a place you can point a virtual camera at.

The one-line version: An AI Location locks a real place's geometry and materials into a reusable model. Time of day, lighting, weather, and how many people are in the room stay variable — you set those per shot.

Why it matters¶

Location is one of the most expensive, least repeatable parts of filmmaking. You book a space for a day, fight the sun, lose access, and can never quite get the same room back for reshoots. AI Locations break that.

Shoot the same place across an entire film. Episode 1's apartment is identical in Episode 8. No continuity drift, no rebuilding a set.

Re-light without re-shooting. The same location renders at golden hour, blue hour, or pitch dark — because it was trained to. You're not stuck with the one look you captured.

Fill the room — or empty it. Generate the café empty for a quiet morning scene or packed for the lunch rush, from one model.

No permits, no travel, no weather delays for reshoots. Once the location exists as a model, it's always available.

How it works¶

The core idea is one rule: geometry constant, everything else variable.

A real location has two kinds of properties. Some never change — the layout, the walls, the materials, the fireplace, the neon sign over the bar. Those are the location's identity. Others change constantly — the time of day, which lamps are on, whether it's raining, how many people are inside. Those are conditions.

An AI Location is trained to learn only the identity and treat every condition as adjustable. That's why one model can render the same room at dawn or midnight: it was deliberately shown both, so it never bakes in a single look.

To make that work, capture is spread across deliberately diverse conditions:

• Time of day — pre-dawn, morning, midday, afternoon, golden hour, blue hour, night
• Lighting states — all lights on, single lamp, TV glow, edge-lit dark, colored wash
• Weather — clear, overcast, rain, post-rain (for anything with a visible exterior)
• Occupancy — empty, a few people, a crowd

Every variant is shot from the same anchor camera angles, so the model lines the geometry up across all of them and learns the space as constant while the lighting shifts. A model trained only on bright midday photos will force every shot to look bright and midday — diversity is what makes it flexible.

Four location types¶

Not every place varies the same way, so AI Locations come in four types. The type tells a buyer exactly what the model can do.

The type isn't a label — it's a promise. A Type C nightclub is sold on its lighting states (peak color wash, empty-before-open, edge-lit dark), because that's what a filmmaker buying it actually needs.

How AI Locations fit with the other models¶

An AI Location is one of six model types on Lookhouse, and they're designed to stack. A single shot typically combines a few:

• Character (the actor's identity) +
• Action (the body motion) +
• AI Location (the place) +
• Style (the film's visual look)

Locations are reusable — the same place works with any character, any action, any style. Stack 3–4 models per shot for best results; quality starts to degrade past 4–5. The location handles where; the other models handle who, what, and how it looks.

For each audience¶

If you're a Buyer (AI film producer)¶

You're buying a place you don't have to scout, rent, or relight. Before you buy, the listing tells you exactly what you're getting: the location type, a floor plan with camera positions, reference grids showing every supported time-of-day and lighting state, an occupancy range (empty → crowded), and sample renders across that whole matrix. If a model claims golden hour, there's a golden-hour render to prove it.

To use one: load the AI Location alongside your character and style models, then prompt the condition you want — "golden hour, empty, wide establishing from the entryway" or "nightclub peak, color wash, crowded, low angle from the dance floor." The place stays consistent; you direct the light and the crowd.

Know the limits going in (see below) and storyboard around them. A good listing names its own failure modes and the workaround for each.

If you're a Creator (actor, model, location owner)¶

If you own or have access to a distinctive space, it's an asset. Capture it well once and it can sell on the marketplace to filmmakers who need exactly that vibe — a Brooklyn brownstone living room, a smoky jazz bar, a brutalist lobby.

What makes a location sell is variety, not volume. Buyers pay for a place they can use in many scenes, which means your capture has to cover the range: different times of day, different lighting states, empty and full. A space shot only at noon, only empty, is worth far less than the same space captured across its full range. Lookhouse's capture guide walks you through exactly what to shoot — you provide the footage, we handle the training.

One rule worth knowing upfront: keep identifiable strangers out of your shots. Anyone clearly visible can get bound into the model and bleed into later scenes — and it's a release/privacy problem. People should be background density (turned away, in motion, or in silhouette), never head-on portraits.

If you're a Developer¶

You already train LoRAs — Lookhouse is where you sell location models to filmmakers. The training discipline that makes a location model marketplace-grade:

• Identity via trigger token, conditions via caption. Caption everything that varies (time, lighting state, weather, occupancy, camera angle) so the model learns those as adjustable. Don't caption the invariant architecture — the trigger token carries identity, and re-describing fixed elements dilutes the binding.
• Develop the whole dataset with one preset. Neutral profile, consistent white balance, no creative looks, no HDR tone-mapping. Inconsistent develop becomes part of the location's learned "identity."
• Hold out a validation set (10–15 images across the variation axes) before training. Run a fixed eval prompt set at every checkpoint — including a negative test: drop the trigger and confirm the location collapses to a generic interior. If it still renders, the LoRA leaked into the base weights.
• Balance across states. If the single-lamp render refuses to go dark, the dataset was too bright. Re-balance, don't just train longer.

Ship the model with its documentation pack — type declaration, spatial map, reference sheets, support matrix, and known limitations. That pack is what converts a browsing filmmaker into a buyer.

What it can and can't do¶

Set expectations honestly — it builds trust and saves buyers from mid-project surprises.

Reliably delivers:

• The same recognizable space across time of day, lighting, weather, and occupancy
• Consistent materials — wood, brick, fabric, tile, metal
• The distinctive features that make the place itself — the fireplace, the neon sign, the chandelier
• A plausible, stable layout in single-shot generations

Doesn't fix (these are limits of today's base video models, not of the location model):

• Architecture across cuts — a door on the left wall in one shot may move or vanish in another generated from a different prompt. Pair matching shots with the same seed.
• Window views — the model tends to invent what's outside. Reference-image conditioning helps.
• Mirrors and reflections — often incoherent, especially in motion.
• Long takes — geometry typically holds for ~4–6 seconds before drifting. Plan shorter shots where continuity matters.
• Fine repeated patterns — wallpaper, intricate art, and detailed fabric may simplify or warp.

The fix for most of these is craft, not a better model: split shots, pair seeds, condition on a reference image, and storyboard around the known soft spots.

In short¶

An AI Location turns a real place into a reusable asset: captured once, rendered forever, in any condition you can prompt. Buyers get a location they never have to scout or relight. Creators turn a great space into recurring marketplace income. Developers get a marketplace built for selling exactly this. Geometry stays locked; the light, the weather, and the crowd are yours to direct.

Part of the Lookhouse model library — AI Location is one of six model types you can build, buy, and stack: Character, Action, AI Location, Style, Camera Motion, and Transition.