DeStyle2Style: Scalable Destylization-Driven Data Generation for Artistic Style Transfer

ICLR 2026 Conference Submission14628 Authors

Keywords: Style Transfer, Destylize

Abstract: DeStyle2Style introduces a novel approach to artistic style transfer by reframing it as a data problem. Our key insight is destylization, reversing style transfer by removing stylistic elements from artworks to recover natural, style-reduced counterparts. This yields DeStyle-100K, a large-scale dataset that provides authentic supervision signals by aligning real artistic styles with their underlying content. To build DeStyle-100K, we develop DestyleNet, a text-guided destylization model that reconstructs style-reduced natural images, and DestyleCoT-Filter, a multi-stage evaluation model that employs Chain-of-Thought reasoning to automatically discard low-quality pairs while ensuring content fidelity and style accuracy. Furthermore, we introduce BCS-Bench, a benchmark with balanced stylistic diversity and content generality for systematic evaluation of style transfer methods. Our results demonstrate that scalable data generation via destylization offers a reliable supervision paradigm, effectively addressing the fundamental challenge of lacking ground-truth data in artistic style transfer.

Primary Area: datasets and benchmarks

Submission Number: 14628