Reference¶
The shared vocabulary and rules behind Lookhouse — written once here and linked from every Creator, Buyer, and Developer guide. If a term, base model, or license tier shows up in another page, this is where it's defined.
Start with What is a LoRA? if you're new, jump to the Compatibility matrix to check what runs where, or use the Glossary for a quick definition.
What is a LoRA?¶
A base model (SDXL, Flux, and the like) already knows almost everything — faces, fabric, sunlight, motion, depth. What it doesn't know is your particular person, place, look, or move. Retraining a billion-parameter model to learn one of those would cost a fortune.
A LoRA (Low-Rank Adaptation) is the shortcut. It freezes the base model and trains a tiny set of new weights that nudge it toward one specific thing — a thin overlay clipped onto a finished engine. The engine does the heavy lifting; the overlay steers it. Because the thing it adds is small, the file is small, and you control how hard it lands with a strength dial at generation time.
The payoff is consistency. Ask a raw model for "the same person" or "the same anime style" across ten shots and you get ten variations the audience can feel. A LoRA holds it steady, frame after frame.
The six LoRA types¶
The same set whether you're buying or selling. Each links to its full guide.
| Type | What it captures | Trained on |
|---|---|---|
| AI Character | A person who reads as themselves across prompts — looks and performance | Images + video |
| AI Action | A motion pattern — walks, gestures, sport-specific moves, with correct weight and timing | Video |
| AI Location | A place you can re-light and re-crowd — time of day, angles, coverage | Images |
| AI Anime | A specific rendering style — line weight, palette, shading — held steady across a production | Images |
| AI Camera Motion | A camera move within a shot — dollies, push-ins, handheld, gimbal | Video |
| AI Transition | A move between two shots — cross-fades, whip-morphs, day-to-night reveals | Video (pairs) |
Camera Motion moves the camera within a single shot; a Transition moves between two shots. They pair well but solve different problems.
Base models & compatibility¶
Every LoRA is built for a specific base model — the foundation it runs on, whether a still-image model like SDXL or Flux, or a video model such as Wan. Matching matters: a LoRA delivers its quality only on the base it was built for.
You don't need a master list — compatibility is shown on every listing, and the right version reaches you automatically:
- Ordering a creator's character? You place the order and we build it for the base model your project uses, so it arrives ready for your pipeline.
- Buying a developer's upload? The listing shows exactly the version(s) the developer published — upload a Wan build, buyers see the Wan version. Pick the one that matches your stack.
- One subject, several bases. A single listing can carry the same subject trained for multiple base models; choose yours at checkout and get the matching file.
Rank, role, and how they map¶
Lookhouse organizes LoRAs by role — how large a part the LoRA plays in a shot — and role is read from the rank of the file. Because each type does a different amount of work, the same role maps to a different rank per type.
| Role | Character | Location | Transition |
|---|---|---|---|
| Background | 8–16 | 16 | 16 (Subtle/Effect) |
| Supporting | 16 | 32 | 16 DoRA / 32 (Standard Hero) |
| Hero | 32 | 32 DoRA / 64 | 32 DoRA / 64 (Complex Hero) |
| Marquee | 32 DoRA / 64 | 64 DoRA | 64 (Pack) |
Three rules of thumb: DoRA at rank r ≈ plain LoRA at rank 2r with a smaller footprint (use it for premium tiers); rank 128 is rarely right — split into bundled lower-rank LoRAs instead; and test composition before listing, since buyers stack character + location + transition together.
Glossary¶
Base / checkpoint — the foundation model a LoRA is built to run on (e.g. an SDXL- or Flux-family model). A LoRA is matched to a base; the right one matters for quality.
LoRA (Low-Rank Adaptation) — a small set of trained weights that bolts onto a base model and nudges it toward one specific identity, place, look, or move.
DoRA — a LoRA variant that delivers comparable quality at roughly half the rank with a smaller weight footprint, so it composes more cleanly when stacked.
Rank — how much "room" a LoRA has to store detail. Higher rank holds 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 result; ease it down to let it blend.
Trigger word — a unique token (like a stage name) dropped in the prompt to call the LoRA into the scene.
Role — the scale of the part a LoRA plays in a shot: Background, Supporting, Hero, or Marquee.
Usage — how the buyer intends to use the output: Personal, Creator, Commercial, or Broadcast.
Studio — the Lookhouse web application where creators train LoRAs and buyers generate videos. Always capitalized.
Licensing terms¶
A license is set along two independent axes. The buyer picks one from each, and the combination sets the price.
Role — the scale of the part:
| Role | Description |
|---|---|
| Background | Extras / crowd |
| Supporting | Secondary cast |
| Hero | Lead / close-up |
| Marquee | Top-tier packs |
Usage — what the output is for:
| Usage | Label | Description |
|---|---|---|
| Personal | Hobby Use | Hobbyists, social posts, personal portfolios — no money changes hands |
| Creator | Indie Work | Solo creators and small studios doing paid client work under ~$50k revenue |
| Commercial | Agency / Brand | Agencies and brands running funded campaigns, paid social, paid digital |
| Broadcast | TV / Film | Linear TV, streaming originals, theatrical — the high-reach tier |
Because the axes are independent, the same file can be licensed for a hobby project or a broadcast production without separate listings. Full detail lives in How licensing works.
Content policy¶
Stub — needs the actual marketplace rules. What may and may not be published: consent and likeness requirements for Character LoRAs, prohibited subjects, and how violations are handled.