How to stop your generations looking like everyone else's

You can usually tell which model made an image. Not because it is bad — because it has a house look: a particular gloss, a particular way with skin and light, a particular idea of what "cinematic" means. Everybody is prompting the same few models, so everybody's output rhymes.

Prompting harder does not fix this. You can push a model some distance with words, but the words have to be re-argued every single generation, and the look drifts anyway across a series — which is precisely where a recognisable channel is made or lost.

The fix is a style: a small trained adapter that rides along with the model and bends everything it makes towards one look. There are three ways to get one, and only the last involves training anything.

Route 1 — Take one off the shelf

The ready-made shelf is a set of styles you can use immediately: photography looks, illustration looks, and video looks for generated clips. You pick one, generate, and pay nothing beyond the generation you were making anyway.

These are public adapters, and the shelf tells you what you are standing on: the source it came from and the licence it carries are on the tile, because anyone selling what they generate is entitled to see the licence without hunting for it. Where an author has stated no licence, it says that too — an author who states no licence has granted nothing, which is a different situation from a permissive one and should not look the same.

Most of these styles only work when a particular trigger phrase appears in the prompt. That is handled for you and injected at generation time — it is never glued onto your stored prompt, so reusing an old prompt later does not hand you back somebody else's magic word.

Start here. It costs one generation to find out whether a look is what you wanted, and that is a much better trade than the same discovery after a training run.

Route 2 — Import one you found

If you have found a style elsewhere — the public model hubs are full of them — paste its link and it is checked before anything is saved.

The check that matters is compatibility. A style is trained against a specific base model, and one trained for an image model simply cannot be loaded by a video model: it is an error, not a bad result. Because the person pasting a link has no way to know that, the refusal names the reason rather than failing vaguely later.

What comes back before you commit is the model it fits, the licence, and the trigger word. If the author states no licence, you are told plainly, along with the fact that this does not mean it is free to use commercially.

Two things worth knowing about public styles generally: download counts measure enthusiasm rather than quality on your prompts, and a style trained on a named real person is a legal problem regardless of how good it looks. Ours are kept off the shelf entirely for that reason.

Route 3 — Train your own

This is the one that produces a look nobody else has, and it is the only route that starts from your pictures.

Training takes about a dozen images and learns one of three things, and the difference matters more than it sounds:

• A look — pictures that share a style, so everything you generate afterwards carries it.
• A subject — a product or a person, so the same thing can appear in scenes you never shot.
• A video look — trained on your own clips, for video you generate from scratch.

Choose the material, name it after the thing rather than after yourself ("House look", not "My style"), and start. The price in tokens is on the button before you press it, it runs on our side so you can close the tab, and if a run fails the tokens come back automatically.

Afterwards there is nothing to type: the style is applied for you wherever you generate, with a strength slider under it.

The dozen pictures that decide everything

This is where trained styles succeed or fail, and it is not about quantity.

Twelve pictures that share one look teach more than forty that share nothing. A mixed set is the single most common reason a style comes out weak — the model averages what it was shown, and the average of five unrelated aesthetics is mud.

So: same palette, same lighting philosophy, same level of finish. Vary the subject as much as you like — that is what teaches it the look is separable from the content — but do not vary the look. If you are training on your own illustration, do not include the three pieces from the year you were experimenting.

For a subject rather than a look, invert it: same subject, varied angles, lighting and backgrounds.

When it comes out too faint

Almost always the strength slider, which costs nothing to move. Turn it up and the look asserts itself.

If a result at full strength still looks untouched, retrain from the style card — same files, one click, no re-picking. And if that also disappoints, the honest answer is usually the training set: go back and cut it down to the pictures that genuinely agree with each other.

Do not leave every style pinned at maximum out of habit. A style at full strength will happily flatten the composition you asked for, and the sweet spot for most looks is short of the top.

What a style will not do

It bends the look. It does not fix the picture.

• Composition, anatomy and hands are the model's job, not the style's. A style trained on beautiful photographs will not stop a model putting six fingers on a hand.
• It is model-specific. A style belongs to the base it was trained against; switching to a different engine means picking a style for that engine.
• Image looks and video looks are different things. A look trained on stills applies to generated stills; carrying a feel into generated video needs a style trained on clips.
• It is not a brand kit. Fonts, colours and logo are the text and grading side of consistency, and they do more for a finished video than any adapter.

Likenesses, briefly and seriously

Training on a person is allowed when it is you, or when you have their explicit written permission — you are asked which before training, and the answer is recorded with the wording you agreed to.

This is not a formality. Training on somebody's likeness without permission is unlawful in many places, and the account that trained it is responsible for what happens next. It is also why no likeness style ever appears on the ready-made shelf, whatever its licence or popularity.

Your own trained styles stay private to your workspace, are never offered to anyone else, and the trained file is deleted permanently when you delete the style.

The order that works

1. Try a ready-made style first — one generation tells you whether the look is right.
2. Import a public one if you have found something closer, and read the licence it comes back with.
3. Train your own when you need a look that is genuinely yours.
4. Curate the dozen pictures ruthlessly: one look, varied subjects.
5. Move the strength slider before you conclude anything is wrong.
6. Retrain from the card if full strength is still faint.
7. Keep composition, grading and typography doing their own jobs — a style is one layer of a look, not all of it.