Specialty Digest

Global English editionLearners’ Study Dashboard — in development
The Knowledge Library for All

Style Transfer

nounGenerative AIalso neural style transfer, artistic style transfer, image style transfer

In one line

Style transfer is an AI technique that re-renders the content of one image, video or text in the artistic style of another while keeping its subject recognizable.

Style transfer is an artificial intelligence technique that re-renders the content of one image in the artistic style of another. A photograph of a harbor can be turned into something that looks like a painting by Van Gogh or a woodcut print, with the buildings and boats still recognizable but with new brushstrokes, colors and textures. The best-known version is neural style transfer, introduced in 2015, which uses a convolutional neural network to separate and recombine the content of one picture and the style of another. Style transfer is an early and influential example of generative AI applied to art, and it is part of the wider field of image generation and computer vision.

History of style transfer

Before deep learning, researchers in computer graphics studied non-photorealistic rendering, which made pictures look like paintings or drawings, and texture synthesis, which extended a sample texture over a surface. In 2015, Leon Gatys, Alexander Ecker and Matthias Bethge published A Neural Algorithm of Artistic Style, showing that a network trained for object recognition, VGG, could be used to combine the content of a photograph with the style of a painting. The results spread widely, and apps such as Prisma brought the effect to phones in 2016. Justin Johnson and colleagues showed how to train a feed-forward network for a single style that applies it in one pass, thousands of times faster. Later work allowed arbitrary styles through adaptive instance normalization, introduced in 2017, and used generative adversarial networks such as CycleGAN to translate images between domains without paired examples. More recently, diffusion models and text prompts allow style changes with words such as in the style of a watercolor.

How neural style transfer works

The original method treats the task as an optimization problem.

  1. Feature extraction: a pretrained convolutional network processes the content image, the style image and a candidate output image, which begins as noise or a copy of the content image.
  2. Content loss: the candidate’s activations in a deeper layer are compared with the content image’s, so that objects and layout are preserved.
  3. Style loss: at several layers, the system computes Gram matrices, which record how strongly different feature channels occur together, capturing textures, colors and brushstroke patterns, and compares them with those of the style image.
  4. Optimization: gradient descent adjusts the pixels of the candidate image to reduce a weighted sum of the content and style losses. See gradient descent.

By changing the weight of the style loss, the user can make the result more faithful to the photo or more painterly. This slow process takes seconds to minutes per image.

Fast and arbitrary style transfer

Feed-forward approaches train a network to produce the stylized image directly, using the same losses during training but not during use. A network for one style can process a video in real time. Arbitrary style transfer methods adjust statistics of the content features, such as mean and variance, to match the style image, so a single model can handle new styles without retraining. These methods trade some quality for speed and flexibility.

Types and related tasks

  • Photo to painting: applying artistic styles to photographs.
  • Video style transfer: stylizing each frame while keeping motion smooth, which requires temporal consistency.
  • Photorealistic style transfer: changing color and lighting, for example making a photo look like it was taken at sunset, while keeping realism.
  • Domain translation: CycleGAN and pix2pix transform images from one domain to another, such as horses to zebras, or sketches to photographs.
  • Text style transfer: changing the tone or style of writing, for example formal to casual, using text generation models.
  • Voice and music style transfer: applying the character of one voice or musical genre to another recording.

Applications

Style transfer appears in photo and video filters, in design tools for creating textures and backgrounds, in games and films for stylized looks and in education to explore art history. Artists use it as part of mixed workflows, designers generate variations of a motif, and researchers use it to adapt synthetic images to look like real ones, which helps train other models. Companies use style transfer for consistent brand imagery.

Strengths and limitations

The approach is simple, visually striking and runs fast once trained. However, it tends to copy textures and colors without understanding the artistic intent behind them, so results can look like a filter. Fine detail such as faces can become distorted, and the result depends heavily on the choice of style image. Video can flicker if frames are processed independently. The method works best on images with clear shapes and does not provide detailed control over where styles apply, although masks can help.

Copyright and ethical issues

Style itself is generally not protected by copyright in many jurisdictions, but a specific work is, and the use of a living artist’s recognizable style raises ethical and sometimes legal concerns. Artists have complained about imitation of their styles without credit or payment. Users should be careful when using a copyrighted painting as the style input, particularly for commercial work, check the license of the tool and the images, credit the original artist where appropriate and avoid presenting generated pieces as the work of the person whose style was copied.

A worked example

A teacher wants an engaging poster about a local harbor. She takes a photograph and chooses a public domain Japanese woodblock print as the style image. Using a style transfer tool, she sets the content weight high so that boats remain clear and the style weight moderate so that the texture appears in the water and sky. After trying three versions, she picks one, prints it with a caption naming the print that inspired the style and notes that the image was produced with software.

FAQs

Is style transfer the same as a photo filter? Simple filters adjust color and contrast, while style transfer uses a neural network to change textures and brushstrokes in a content-aware way.

Can style transfer copy any artist’s style? Technically it can imitate many styles, but ethical and legal questions apply, especially for living artists.

Does style transfer use the original painting’s pixels? No. It learns statistical patterns of colors and textures and applies them to the new image.

Key facts

  • Neural style transfer was introduced in 2015 by Leon Gatys, Alexander Ecker and Matthias Bethge, using features from a pretrained convolutional network.1
  • The method separates content, taken from deep layers, from style, captured by correlations between feature maps called Gram matrices.1
  • Later feed-forward networks made style transfer fast enough for real-time and mobile use.2

Related concepts

Quick checkHow does neural style transfer separate content from style?Show answer

It uses a pretrained convolutional network, taking content from the feature activations of deeper layers and style from the statistics of correlations between feature maps across layers, then optimizes a new image to match both.

Sources

  1. Gatys, Ecker and Bethge, A Neural Algorithm of Artistic Style, https://arxiv.org/abs/1508.06576
  2. Johnson, Alahi and Fei-Fei, Perceptual Losses for Real-Time Style Transfer and Super-Resolution, https://arxiv.org/abs/1603.08155

Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue