A saved prompt paired with a model that suits it. It is not a trained vector or a filter applied after generation. Because it is a prompt, you can layer your own description on top of a preset rather than being limited to a single strength value.

What styles are available?

Fifteen named video styles under Video Tools: Ghibli style, Disney Pixar, claymation, Lego style, pixel art, pencil drawing, vector art, vector illustration, artistic, and several anime variants including classic, modern, 3D, simplified 3D and stylized 3D. Image Tools carries a separate set of presets for jobs such as photo restoration, watermark removal, vectorising, stickers, retouching and text removal.

How do I describe a style that has no preset?

Be concrete about what produces the look rather than naming an aesthetic. "Thick black outlines, flat colour fills, limited palette, hard shadows" works better than "comic book style", because it tells the model what to do instead of what to reference.

Can I save a custom style to reuse?

There is no per-workspace style library and no capacity limit on one, because nothing is being stored as a style. Save your prompt wording yourself and reuse it. Claims about a set number of custom style vectors per workspace describe a feature that does not exist.

What Are AI Styles, and How Do They Work

Q: How do I set style strength?

You can't, because there is no style strength parameter. Across all 37 video generation models the controls are prompt, duration, aspect ratio and resolution, plus reference images, a negative prompt or first and last frames on some models. No entry accepts a style weight. To get more or less of a look, change the wording or supply a reference image.

Q: Why does the same style look different on different subjects?

Because it is not a filter being applied on top. It is an instruction asking the model to generate something a certain way, and the subject is part of that request. A style that suits a landscape may read differently on a portrait.

Q: How do I keep a style consistent across several images?

Reuse the same wording exactly and change only the subject and action. Rewriting the style description each time invites the model to reinterpret the whole look. For a face or product that must stay itself, a reference image does what no wording can, and for a sequence the Long Video Generator chains segments so each continues from the last.

A style is a saved prompt, not a dial.

That single fact explains most of the confusion around this. There is no style strength slider, no vector you scale from 0.2 to 1.0, no library of trained aesthetic embeddings you manage. What exists is a preset: a carefully written instruction, attached to a model that suits it, sitting behind a button so you do not have to write it again.

Once you know that, styles stop being mysterious and start being editable.

What you actually get

Two kinds of preset, in two places.

Video styles. Fifteen named looks under Video Tools: Ghibli style, Disney Pixar, claymation, Lego style, pixel art, pencil drawing, vector art and vector illustration, plus several anime variants including classic, modern and 3D. Each is its own page.

Image tool presets. A separate set under Image Tools covering jobs rather than aesthetics: restore an old photo, remove a watermark, turn a photo into a vector, make a sticker, retouch, remove text, build a 3D character.

The difference matters. Video styles change how something looks. Image tool presets change what something is.

Why "there is no strength slider" is good news

If a style were a trained vector with a strength value, your only lever would be that number. Turn it up, get more of a fixed thing. Turn it down, get less.

Because a style is a prompt, you have every lever a prompt has. You can layer your own description on top, ask for the look applied to a specific subject, or specify lighting the preset never mentioned. The preset is a starting point rather than a mode you enter.

This is also why the same style behaves differently on different subjects. It is not applying a filter, it is asking a model to generate something a certain way, and the subject is part of that request.

How to actually get a look you want

Three approaches, in order of how much control they give you.

Use a preset. Fastest, and correct when the look you want has a name. Ghibli, claymation, pixel art. Go to the page, supply your input, generate.

Describe it yourself. More control, and necessary when your look has no name. Be concrete about the things that produce a style rather than naming an aesthetic: line weight, colour palette, lighting direction, texture, level of detail. "Thick black outlines, flat colour fills, limited palette, hard shadows" gets you further than "comic book style", because it tells the model what to do instead of what to reference.

Show it. The most reliable of the three. Models that accept a reference image will match a look far more closely than any description of it, because a picture is not a lossy description of itself. Reference support exists across most image models and thirteen video models.

If you have an example of the look you want, use it. Describing an image you are holding is doing extra work to get a worse result.

Consistency across a set

The usual reason people go looking for a style strength value is that shot three does not match shot one.

The fix is not a number. It is reusing the same instruction. Keep your style wording byte-identical across generations and change only the subject and action. Rewriting the description each time invites the model to reinterpret the whole look.

For a face or product that must stay itself, a reference image does what no wording can. For a sequence that must hold together shot to shot, the Long Video Generator chains segments so each continues from the last.

What is not real

Stated plainly, because these claims are everywhere.

There is no style strength parameter. Across all 37 video generation models the controls are prompt, duration, aspect ratio and resolution, plus reference images, a negative prompt or first and last frames on some models. Image models are similar. No entry accepts a style weight.

There is no per-workspace style library with a capacity limit. You cannot save a custom style vector, and there is no cap on how many to keep, because there is nothing being stored.

Style adherence is not scored. No cosine similarity, no CLIP metric surfaced anywhere. If an article quotes a style-match figure to two decimal places, it made it up.

Picking a model for a look

Style is mostly prompt, but the model still matters.

Illustration and graphic looks suit models strong on clean shapes and typography. Ideogram V4 and Recraft are built around that kind of output.

Photographic looks suit models tuned for realism, such as Cosmos 3 Super or MAI-Image-2.5.

Editing an existing image into a style is a different job from generating one. FLUX Kontext Pro handles broad transformation, GPT-Image 2.0 Edit does precise single-region work with detail preservation.

Video styles run through the preset pages, which already pair each look with a model that suits it — that pairing is a large part of what you are getting from a preset.

The full catalog is at Models, and the app quotes each generation before you run it.

The short version

A style is wording plus a suitable model. Presets bundle both so you can skip the writing.

When a preset gets you close but not there, do not go looking for a slider. Add your own description on top, or supply a reference image of the thing you actually want. Those are the two real levers, and between them they cover almost everything people are trying to do when they search for style controls.