AI Action¶

Motion that holds up under prompt drift — walks, gestures, fight beats, sport-specific moves.

One of the six LoRA types on Lookhouse. A Character LoRA makes a person look right. An AI Action model makes a body move right. This guide explains what it does, how it works, why it's unique, and what a good sample looks like — for creators selling motion, buyers licensing it, and developers wiring it into a pipeline.

The short version¶

Most LoRAs teach a diffusion model about appearance — a face, a place, a drawing style. AI Action teaches the model about movement: the way a particular body carries weight, plants a foot, throws a punch, swings a racket, or signs its name in the air.

It's the one type on Lookhouse that isn't really about how a frame looks. It's about how frames change over time — the thing that makes AI footage read as real or fake the instant a body starts moving.

For creators: your craft — how you move — becomes a sellable asset, not just your face.
For buyers: you get motion that survives prompt changes, so a character keeps moving like that fighter, that dancer, that athlete across every shot.
For developers: it's a motion-conditioning LoRA in the submission pipeline, distinct from identity and style LoRAs, and stackable with them.

What an AI Action model actually is¶

A LoRA (Low-Rank Adaptation) is a small set of trained weights that bolts onto a large base model and nudges it toward something specific. An AI Action LoRA is trained on video — sequences of frames — rather than still images, so the thing it learns is a motion pattern rather than a look.

Concretely, it captures:

Kinematics — the path a body and its limbs trace through space: stride length, arm arc, hip rotation, follow-through.
Timing and weight — acceleration, the pause at the top of a jump, the snap of a jab, how impact settles. This is what animators call "weight," and it's the hardest thing for generic models to fake.
Style of motion — two people can throw the same punch and look nothing alike. AI Action encodes that signature, not a generic version of the move.

It does not primarily encode identity (that's a Character LoRA) or rendering style (that's Anime Style) or camera behaviour (that's Camera Motion). It rides on top of those.

How it works, end to end¶

1. Capture. The creator records video of the move — clean, repeated, from useful angles. Footage, not photos. (See What a good sample looks like below.)

2. Train in Studio. Lookhouse extracts the motion across frames and fits a low-rank adapter against a video-capable base model. The model learns the temporal pattern — how the pose evolves frame to frame — rather than memorising any single frame.

3. List and license. The trained .safetensors file goes into the marketplace with a sample clip, a recommended base model, and a licence tier.

4. Use. A buyer loads the AI Action LoRA in their stack, prompts a scene, and the generated subject moves the way the model was trained to move — usually stacked with a Character LoRA so a specific person performs a specific motion.

Character LoRA   →  who it is        (identity / likeness)
AI Action LoRA   →  how they move    (motion / performance)   ← this guide
Location LoRA    →  where it happens (set / lighting)
Camera Motion    →  how it's shot    (dolly, push-in, handheld)

Stack them and you've cast a person, given them a performance, put them on a set, and handed the camera a move — each piece licensed separately.

What's special about it — and why no other LoRA type covers it¶

This is the key question, so it gets a direct answer.

Every other LoRA type on Lookhouse is fundamentally about a single frame. A Character, Location, or Anime Style LoRA can be judged from one still image — does the face match, is the lighting right, is the line work clean. You could turn off video entirely and still evaluate them.

AI Action cannot be judged from a still. Its entire value lives between frames — in timing, weight, and trajectory. A walk that looks perfect frozen can read as a glide, a moonwalk, or a limp the moment it plays. That temporal dimension is exactly what generic video models get wrong, and it's what no appearance-based LoRA can fix.

So the thing AI Action provides that nothing else does is temporal consistency of a specific motion:

A Character LoRA keeps a person looking like themselves. It says nothing about how they move — drop them into a fight scene and the base model invents the choreography, usually with that floaty, weightless "AI motion" tell.
A Camera Motion LoRA moves the camera. AI Action moves the subject. A push-in on a stiff, generic walk is still a bad walk.
Anime Style / Location LoRAs change rendering and setting. Motion isn't in their vocabulary at all.

Put plainly: AI Action is the only LoRA type that owns the time axis. It's the difference between a character who looks like a boxer and a character who fights like one.

Why it matters¶

It fixes the most visible failure in AI video. Audiences forgive a lot in a still frame and almost nothing in motion. Wrong weight, sliding feet, rubbery follow-through — these are the tells that scream "AI." A motion-trained model is the most direct lever on whether a clip is usable.

It makes performance ownable. Until now, a creator could sell their likeness. AI Action lets them sell their craft — the trained, repeatable way a stunt performer, dancer, athlete, or signer actually moves. That's a new, defensible asset class.

It's reusable production value. A move captured once can be licensed across unlimited projects, restyled, re-cast onto different characters, and dropped into any compatible scene — without re-shooting choreography every time.

It's prompt-stable. The whole tagline — motion that holds up under prompt drift — is the point. Change the wardrobe, the set, the lighting, the camera, and the motion stays put. That stability is what makes a body of shots feel like one performance instead of a dozen near-misses.

What each reader gets¶

For creators¶

Your face is one asset; how you move is another, and AI Action is how you sell it. A fight choreographer, a dancer, a martial artist, a tennis player, a sign-language performer — each has a motion signature that buyers can't get anywhere else. You record the move once in Studio, list it, and earn every time it's licensed. Motion models are especially defensible because they're hard to capture well and hard to fake, which keeps them scarce.

See What you'll actually earn on Lookhouse for ranges and payout cadence by LoRA type.

For buyers¶

You're buying the one thing the base model is worst at: believable, consistent motion. License an AI Action model, stack it on the character you've cast, and that character performs the move correctly across every shot — no re-rolling until the feet stop sliding. Check the licence tier (Personal / Commercial / Ad-Broadcast) before you buy, the same as any LoRA. Confirm the recommended base model matches your stack so the motion transfers cleanly.

For developers¶

AI Action enters the LoRA submission pipeline as a motion-conditioning adapter — same envelope as other types (file, metadata, sample, base-model target, licence), but trained on temporal data and intended to compose with identity and style LoRAs at inference. Expose it in your integration as a stackable layer, surface its recommended base model and frame-rate assumptions, and don't collapse it into the Character slot — they're orthogonal and customers will stack both.

See the API reference, LoRA submission, and Webhooks pages for contract details.

What a good training sample looks like¶

Because the model learns motion, the footage matters far more than the frame.

Do:

Shoot video, not stills — the temporal signal is the whole product.
Isolate the move. One clean action per take, repeated several times, with a neutral entry and exit.
Keep the subject sharp and fully in frame — no cropped limbs mid-motion; the model needs the full kinematic path.
Hold the camera steady (or move it deliberately and consistently) so the model doesn't confuse camera motion with body motion.
Use even, consistent lighting and an uncluttered background so the body is the clearest thing in the shot.
Capture a couple of useful angles if the move reads differently from the side vs. the front.
Keep frame rate consistent across takes; note it for buyers.

Avoid:

Motion blur so heavy the limbs smear into the background.
Baggy clothing or props that hide the joints doing the work.
Cuts, speed ramps, or edits inside a take — the model learns the artefacts, not the motion.
Mixing several different moves into one dataset unless you intend a multi-move model.

A good rule: if a human animator couldn't tell what the move is and how the weight shifts from your footage, the model can't either.

Prompting tips for buyers¶

Stack, don't substitute. AI Action + Character is the standard combo — identity from one, motion from the other. Set sensible weights so neither overpowers the other.
Describe the scene, not the move. The move comes from the LoRA; let your prompt handle wardrobe, set, framing, and mood.
Match the base model. Motion transfers best on the base model the creator trained against — it's in the listing.
Mind the camera. If you also want a camera move, add a Camera Motion LoRA rather than trying to prompt the camera into the action model.

FAQ¶

Is this the same as a Character LoRA? No — and that's the point. Character captures who (a face, a likeness). AI Action captures how they move. They're built to stack: cast the person, then give them the performance.

Can I judge the model from the thumbnail? Never trust a still for AI Action. Always play the sample clip — the value is entirely in the motion, which a frozen frame can't show.

Why not just prompt the motion in words? Text prompts are weak at timing and weight — the exact things that make motion read as real. A model trained on real footage encodes what words can't describe.

Does it lock me to one character or look? No. The motion is portable. The same AI Action model can drive different characters, wardrobes, sets, and styles, because it only owns the movement.

What does it need to run? A video-capable base model — the listing names the recommended one. Match it for the cleanest motion transfer.

In one line¶

If a Character LoRA is who's on screen, AI Action is what they do with their body — the only Lookhouse model type that owns motion through time, and the most direct fix for the thing that makes AI video look fake.

Related: What is a LoRA, in plain English · Character LoRA · Camera Motion LoRA · What you'll actually earn on Lookhouse · How licences work.