Style transfer
last edited
Style transfer re-renders an image in the style of another. The 2015 neural style-transfer papers, which applied "Starry Night" to photographs, were many people's first contact with AI art. The practice has since been absorbed into the diffusion toolkit and is far more controllable.
Methods
The basic approach is img2img with style-laden prompts. The source supplies composition, the prompt supplies the target look, and denoise strength sets how far the restyle goes.
IP-Adapter conditions generation on an example image. Its paper states that the image prompt can also work with a text prompt for multimodal generation. It pairs with ControlNet, which holds the source's structure, so content and style come from different inputs. See combined IP-Adapter + ControlNet workflows.
For a style used repeatedly, a style LoRA trained on example images makes the look portable, strength-adjustable and stackable.
Content-preserving models
Dedicated models now target content-preserving style transfer on top of image-editing backbones. QwenStyle V1 appeared in January 2026. It is described as the first content-preserving style transfer model trained on an image-editing model.
The TeleStyle report, code and model were released on 28 January 2026. A second version followed in June 2026. It supports artistic content reference via self distillation. A later release added open weights, code and a demo.
Ethics
The technique sits at the center of AI art's ethics arguments, since "in the style of a living artist" is exactly what it makes trivial. The detection and attribution debates and moderation fights trace back to this capability. See core techniques.
Sources
- Neural style transfer paper on arXiv, describing content and style recombination
- IP-Adapter paper on arXiv, describing image prompt conditioning
- QwenStyle paper on arXiv, content-preserving style transfer
- TeleStyle repository on GitHub, with report, code and model
- TeleStyleV2 repository on GitHub, with open weights and demo