We disentangle the abstraction and visual appearance in style transfer, allowing independent learning and transferring of structure and style via Visual Analogy Transfer (VAT).
Here we mix different structure and style to create new styles.
Abstract
Artistic styles often embed abstraction beyond surface appearance, involving deliberate reinterpretation of structure rather than mere changes in texture or color. Conventional style transfer methods typically preserve the input geometry and therefore struggle to capture this deeper abstraction behavior, especially for illustrative and non-photorealistic styles. In this work, we introduce Abstraction in Style (AiS), a generative framework that separates structural abstraction from visual stylization. Given a target image and a small set of style exemplars, AiS first derives an intermediate abstraction proxy that reinterprets the target’s structure in accordance with the abstraction logic exhibited by the style. The proxy captures semantic structure while relaxing geometric fidelity, enabling subsequent stylization to operate on an abstracted representation rather than the original image. In a second stage, the abstraction proxy is rendered to produce the final stylized output, preserving visual coherence with the reference style. Both stages are implemented using a shared image-space analogy, enabling transformations to be learned from visual exemplars without explicit geometric supervision. By decoupling abstraction from appearance and treating abstraction as an explicit, transferable process, AiS supports a wider range of stylistic transformations, improves controllability, and enables more expressive stylization.
How does it work?
Abstraction in Style (AiS) formulates stylized image generation as a process that disentangles abstraction from appearance with explicit two stages:
(1) Structural Abstraction focus on learning and transforming the shapes of input image in the spirit of the reference exemplars, producing an intermediate representation (called Abstraction Proxy) that reflects the abstraction tendencies of the reference.
(2) Visual Stylization then takes the Abstraction Proxy and renders it into a final stylized output with colors, textures, etc.
The two stages are implemented within a shared Visual Analogy Transfer (VAT) A → A′ :: B → B′ framework, where Abstraction-VAT learns the analogy "Backbone → Abstraction Proxy" for structure transformation, while Stylization-VAT learns the corresponding "Proxy → Output" for visual stylization.
In practice, VAT is realized with a diffusion transformer and a lightweight LoRA adapter, enabling the model to infer the missing panel B′ from the visible reference pair and the input B, thereby supporting both structural abstraction and visual stylization within a unified framework.
Beyond Color and Texture, Shapes Changed
Below we show the results for the same input image, distinct abstraction proxies are generated by different A-VATs which are trained on different reference respectively.
Below shows the full pipeline of AiS. For each target image, its style-agnostic hidden backbone is first generated.
Then its abstraction proxy is derived from the backbone with the learned A-VAT, which is then rendered into the final stylized output with the learned S-VAT.
See how their shapes are changed in the abstraction proxy, and how the final stylized output is visually coherent with the reference style, demonstrating the effectiveness of AiS in capturing abstraction beyond color and texture.
Try a Quiz? Find which ones are generated
Below is a quiz: guess which stylized objects in the image are results generated by our method, which are the original artworks.
Click the object to see the answer 😊
Comparison with State-of-the-Art Methods
Existing methods struggle with abstract styles and over-rigidly preserve input structure. Our method excels at capturing nuanced artistic styles while maintaining natural structural variations.
More Examples
Citation
@article{lu2026abstraction,title={Abstraction in Style: Beyond Texture and Color},author={Lu, Min and He, Yuanfeng and Chen, Anthony and He, Jianhuang and Wang, Pu and Cohen-Or, Daniel and Huang, Hui},journal={arXiv preprint arXiv:2603.29924},year={2026},}
