Discrete style transfer: Syntax, technique, meaning

This work proposes discrete style transfer (DST) as an alternative to conventional neural style transfer methods, reframing the operation from seamless blending to part-based substitution. Neural models such as VGG-based style transfer, GANs, and diffusion models privilege continuity, producing visually coherent but cognitively estranging images that erase seams essential to human perception, critique, and fabrication. Building on theoretical insights, the paper argues that seams and discontinuities should be treated as design data rather than flaws. The method developed here subdivides images into adaptive quadtree tiles, encodes them with an autoencoder into latent embeddings, and substitutes parts via nearest-neighbor search. Results demonstrate significant gains in speed and stylistic fidelity compared to alternative image comparison methods and mosaic baselines, while remaining competitive with neural style transfer in visual similarity. Beyond efficiency, discrete style transfer produces modular, interpretable outputs aligned with architectural logics of repetition, assembly, and material agency.