AI Transition¶

One of the six AI model types on Lookhouse. Build a signature move once. Use it between any two shots, forever.

What it is¶

An AI Transition is a trained model of a single, repeatable move — a way of getting from one shot to the next, or of transforming a subject inside one continuous take. A liquid morph, a whip-pan blur, a light-flare wipe, an outfit change that happens in a single breath. The model learns the behavior of the transition, not the footage it's applied to, so you can run the same move between any two clips you give it.

Edit a transition by hand once and you've spent an hour matching frames, masking, and re-timing. Train it as an AI Transition and the move becomes a tool: point it at shot A and shot B and it generates the in-between — the same signature way every time, on content it has never seen.

Technically, it's a LoRA: a small (~200 MB) model file that teaches a base video model one specific thing — here, the grammar of one transition — without retraining the whole model. On Lookhouse, we call it an AI Transition because that's what it does for you: it gives you a reusable move you can drop between shots.

The one-line version: An AI Transition locks the motion and timing of a move into a reusable model. What's on either side of it — the subjects, the scenes, the look — stays variable. You supply the two ends; the model handles the journey between them.

Why it matters¶

Transitions are where AI edits fall apart. Two clips that look great alone can cut together badly — a hard jump, a mushy dissolve, a morph that melts into nonsense. Getting a clean, intentional move between AI shots is slow, fiddly handwork, and it almost never comes out the same twice. AI Transitions fix that.

A signature move, applied everywhere. The same morph, swipe, or transformation across an entire film, series, or channel — so the work reads as one piece, not a pile of clips.

Build once, reuse on anything. The move isn't tied to the footage it was trained on. Drop it between any two shots — different subjects, different scenes, different styles — and it behaves the same.

The hard part, baked in. The frame-matching, the timing, the masking craft that makes a transition feel deliberate — it's trained into the model. You direct where it happens, not how.

Transform without cutting. Change an outfit, swing from day to night, or morph an age inside a single unbroken take — no second clip, no edit point.

How it works¶

The core idea is one rule: the move is constant, the content is variable.

A transition has two kinds of properties. Some never change — the path the motion takes, its speed curve, the way it blurs or melts or wipes, how long it lasts. That's the transition's identity. Others change every time you use it — what shot A is, what shot B is, who's on screen, what the scene looks like. Those are the content.

An AI Transition is trained to learn only the move and treat the content as something you plug in. That's why one model can morph a face into a flower or a city into a desert: it was never shown one specific pair often enough to memorize it — it was shown the behavior across many different pairs, so it learned the move as the constant.

To make that work, training is spread across deliberately diverse content:

• Subjects on each side — faces, objects, environments, full bodies, text, abstract shapes
• Motion direction — left-to-right, push-in, pull-out, rise, fall, rotate
• Speed and timing — fast snaps, slow dissolves, ramped accelerations
• Endpoints — visually similar pairs and deliberately mismatched ones

Every example shows the same transition behavior on different footage, so the model lines the move up across all of them and learns it as constant while the content changes. A model trained only on face-to-face morphs will try to force every shot into a face — diversity in the content is what makes the move general.

Four transition types¶

Not every move works the same way, so AI Transitions come in four types. The type tells a buyer exactly what the model can do — and whether it joins two shots or transforms one.

The type isn't a label — it's a promise. A Type D transformation is sold on what it changes and how cleanly the subject survives the change (the face stays the same person, the outfit swaps believably), because that's what a filmmaker buying it actually needs.

How AI Transitions fit with the other models¶

An AI Transition is one of six model types on Lookhouse, and they're designed to stack. Most models build a shot; a transition connects or transforms shots. A sequence typically uses several:

• Character (the actor's identity) +
• Action (the body motion) +
• AI Location (the place) +
• Style (the film's visual look) +
• AI Transition (the move between or within shots)

Transitions are reusable — the same move works on any character, any location, any style. As with the other models, stack 3–4 per shot for best results; quality starts to degrade past 4–5. The transition handles how you get there; the other models handle who, what, where, and how it looks.

For each audience¶

If you're a Buyer (AI film producer)¶

You're buying a move you don't have to build, mask, or re-time by hand. Before you buy, the listing tells you exactly what you're getting: the transition type, example pairs (A → B) shown across varied content, the duration window the move holds for, the kinds of endpoints it handles cleanly versus the ones it struggles with, and sample renders proving each. If a model claims a clean face-to-face morph, there's a face-to-face morph to prove it.

To use one: load the AI Transition, hand it your two shots — or one shot and a target state for an in-shot transform — and prompt the move. "Liquid wipe, left to right, fast" or "day-to-night transformation on the subject, slow, lights coming up." The content is yours; the move stays consistent.

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, signature-style owner)¶

If you have a move people recognize — a transformation, a transition, a way of changing a look that's become your signature — it's an asset. Capture it well once and it can sell on the marketplace to filmmakers who want exactly that move: a seamless outfit-change reveal, a trademark whip-morph, a day-to-night transformation that keeps the face intact.

What makes a transition sell is range, not volume. Buyers pay for a move they can use on many different shots, which means your examples have to cover the range: different subjects on each side, different speeds, similar and mismatched endpoints. A move shown only on one face, only at one speed, is worth far less than the same move shown across its full range. Lookhouse's capture guide walks you through exactly what to record — you provide the example clips, we handle the training.

One rule worth knowing upfront: keep identifiable strangers out of your example footage. Anyone clearly visible can get bound into the model and bleed into later transitions — and it's a release/privacy problem. People in your examples should be you or cleared subjects, never incidental bystanders captured head-on.

If you're a Developer¶

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

• Move via trigger token, content via caption. Caption everything that varies (the subjects, the scenes, the direction, the speed) so the model learns those as adjustable. Don't caption the move itself — the trigger token carries the transition behavior, and re-describing it in captions dilutes the binding.
• Train on many pairs, not many takes of one pair. The same move across diverse content is what generalizes it. Repeating one A→B pair teaches the model that footage, not the move.
• Hold out a validation set (10–15 pairs across the content axes) before training. Run a fixed eval prompt set at every checkpoint — including a negative test: drop the trigger and confirm the model falls back to a plain cut or dissolve. If the signature move still appears, the LoRA leaked into the base weights.
• Balance across endpoints. If the morph only works on faces and smears on objects, the dataset was too narrow. Re-balance the content, don't just train longer.

Ship the model with its documentation pack — type declaration, example pairs, duration window, support range, 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 move applied to footage it has never seen
• Consistent timing and motion feel from one use to the next
• The distinctive signature that makes the move itself — the melt, the whip, the flare, the reveal
• A clean, intentional join between two reasonably compatible shots

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

• Wildly mismatched endpoints — morphing between shapes with nothing in common tends to go mushy in the middle. Pick pairs that share a silhouette or anchor point.
• Frame-accurate beat sync — the move may need manual re-timing to land exactly on a cut or musical beat.
• Long transitions — coherence holds across the move's natural window (roughly the transition itself); stretch it too long and the middle drifts. Keep the move tight.
• Fine detail through the morph — text, faces mid-blend, and intricate patterns may warp or simplify while the move is happening.
• Bad source clips — a transition can't rescue an unstable or low-quality A or B. Garbage in, garbage out.

The fix for most of these is craft, not a better model: choose compatible endpoints, keep the move short, condition on clean source shots, and storyboard around the known soft spots.

In short¶

An AI Transition turns a signature move into a reusable asset: built once, applied between any two shots forever, on content it has never seen. Buyers get a clean, intentional join they never have to mask or re-time. Creators turn a trademark move into recurring marketplace income. Developers get a marketplace built for selling exactly this. The move stays locked; the shots on either side are yours to direct.

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