OmniStyle2

We present DST-100K, a high-quality dataset for style transfer, constructed through a novel destylization-based pipeline. The goal of destylization is to reverse the stylization process by recovering the underlying natural appearance of artistic images. This formulation transforms the original style image into an authentic supervision signal, enabling the learning of style transfer from real styles with aligned content. To achieve destylization, we design DST, a text-guided destylization model which can remove stylistic features from artistic images to generate style-free natural counterparts guided by content text. Because imperfect destylization would propagate noise into downstream training, we further introduce DST-Filter, a Chain‑of‑Thought, multi‑stage evaluation that jointly measures content preservation and style discrepancy, automatically discarding low‑quality pairs. Leveraging DST‑100K, we build DST-Transfer, a feed-forward style transfer model based on FLUX.1-dev without adding any new modules and handcraft design. Despite its simplicity, DST‑Transfer consistently surpasses state‑of‑the‑art methods in qualitative and quantitative evaluations. Our approach reframes style transfer as a data problem and introduces a reliable supervision paradigm derived directly from authentic artistic styles, which helps address the critical challenge posed by the absence of ground truth data in style transfer tasks.