TeleStyle V2

TeleStyle V2 is a research model for content-preserving style transfer: given a content reference image and a style reference image, it generates an output that keeps the content and adopts the style. It follows TeleStyle V1, which was trained only on photorealistic content with artistic style references. V2 extends this to realistic-and-realistic, realistic-and-stylized, stylized-and-realistic and stylized-and-stylized pairings, using a self-distillation strategy that builds training triplets from V1.

The released variant, TeleStyleV2-QIE-2509-DMD, is fine-tuned from Qwen-Image-Edit-2509 and uses Distribution Matching Distillation. According to the authors, this cuts inference cost by about 10 times and helps retain the base model's general text-guided editing ability and content consistency. It accepts a variable number of reference images: style reference plus a prompt, a single content image for ordinary text-guided editing, or both. A prompt enhancer based on Qwen2.5-VL-7B generates content and style prompts to address reference-order confusion seen in V1.

The authors report quality comparable to gemini-3-pro-image-preview on style transfer and at least on par with Qwen-Image-Edit-2509-DMD on general editing. The project page provides a tech report and BibTeX (arXiv 2606.20709). Code is reported released by a YouTube creator, but no repository, weights link or licence was available in the provided material. The vendor is not stated on the page; the authors are Zhang, Xu, Huang, Zhang and Li.