Van Gogh Style AI Filters: How They Work

A style filter looks like a filter — something laid over the top of your photo. It is not. It is a re-rendering, and understanding that explains why the results vary so much between photos.

What is actually being separated

The useful idea here is old, and predates diffusion models: an image can be thought of as having content — what is in it, where things are, the arrangement of shapes — and style — how it is rendered: stroke, texture, palette, edge quality.

A style transfer keeps the first and replaces the second. Your photo supplies the composition. The style supplies everything about how that composition is drawn.

Which means, unavoidably: fine detail is not content, it is style. The texture of skin, individual hairs, the weave of fabric — those live in the layer being replaced. They do not survive. What survives is the arrangement.

That single fact predicts almost every good and bad result.

What "Van Gogh style" means to a model

The model learned the visual signature of the style as a cluster of qualities:

• Thick, directional strokes with visible texture, following the form of what they describe
• High-contrast complementary colour — the blue-and-yellow relationship most people associate with the late work
• Movement in flat areas. Skies and fields that would be smooth in a photograph get swirl and rhythm
• Outlined forms with heavy, darker edges
• Compressed tonal range. Fewer, bolder values rather than smooth gradients

It is reproducing those qualities. It is not sampling from paintings, and it cannot give you The Starry Night with your house in it — it can give you your house rendered with the qualities that painting shares with the rest of the body of work.

Which photos work

The pattern follows directly from content surviving and detail not:

Restyles well:

• Landscapes with a clear horizon and large tonal areas
• A single subject against an open background
• Architecture — strong geometry, simple planes
• Anything with a bold, readable composition at a glance

Restyles badly:

• Busy scenes with many small elements, which become visual noise
• Fine repeating texture — foliage, crowds, patterned fabric
• Photos whose interest is the detail, because that is what gets replaced
• Small faces, which lose the features that make them recognisable

Before restyling, squint at the photo until the detail disappears. What you can still see is what will survive. If nothing remains, the result will be mush.

Portraits need care

Faces are the hardest case, because recognition depends on exactly the fine structure a painterly style removes.

Two things help:

Make the face large in frame. A head-and-shoulders crop keeps more structure than a full body.

Ask for a lighter touch. "A soft watercolour wash, keeping the facial detail" behaves very differently from "a thick impasto oil painting". Naming what to preserve genuinely works, because it competes with the style instruction rather than leaving it unopposed.

And accept the trade. A heavily stylised portrait is a picture of someone rather than a likeness. That is fine when it is the point — it is not fine when you wanted a headshot, which is a different tool entirely.

Beyond Van Gogh

The same mechanics apply to every style, and the differences are practical:

• Watercolour — soft edges, blooms, paper texture, light tonal range. Forgiving on faces. See turning a photo into a watercolour.
• Oil — visible strokes, heavier detail loss than watercolour
• Anime and cartoon — flat colour, hard outlines, simplified features. A different kind of transformation, covered in turning a selfie into anime art
• Pencil and ink — tone becomes hatching, colour disappears entirely

Choose based on how much detail your photo can afford to lose.

A workable approach

1. Squint at the photo. If the composition does not read, pick another.
2. Name the style and what to preserve.
3. Generate two or three times — results vary and one is usually clearly better.
4. Check faces for drifted anatomy, which is the first thing to go when style dominates.

Restyling is one of the four tools in Kitana's browser studio, so it costs nothing to see how your photo behaves. The full set of thirteen photo tools is in the apps.