Key points
- The method separates and recombines image content and artwork style using deep VGG feature statistics.
- It lets artists repaint an image in any reference style without paired training data, the conceptual root of neural stylisation.
- It founded the neural-style-transfer subfield and is in nearly every creative-AI curriculum.
Summary
Gatys, Ecker and Bethge (2015/2016) showed that the Gram matrices of convolutional feature maps from a pretrained VGG network encode visual style, and that content and style can be independently recombined by optimising an image against both a content loss and a style loss. The result is an input photograph rendered in the visual idiom of any reference artwork without paired training examples. The paper founded the neural-style-transfer subfield, triggered a wave of commercial style-transfer apps, and remains the conceptual root of style conditioning in diffusion-era tools such as ControlNet and LoRA. It appears in nearly every creative-AI curriculum as the starting point for discussions of neural stylisation and AI-assisted artwork.
Related items
Source
Source: CVPR 2016 (IEEE/CVF) ↗ (Research)
Cite this item
Gatys, L. A., Ecker, A. S. & Bethge, M. (2016). ‘A Neural Algorithm of Artistic Style (neural style transfer)’, CVPR 2016 (IEEE/CVF). Available at: https://arxiv.org/abs/1508.06576
Your reference manager can also read this page directly: with the Zotero (or Mendeley) browser connector installed, save it straight to your library. Whole-collection exports: RIS, BibTeX, CSL-JSON.