TeleStyle

TeleStyle is an open-source style transfer model that keeps the source content intact, released by the Institute of Artificial Intelligence (TeleAI) at China Telecom. The team behind it includes Shiwen Zhang, Xiaoyan Yang, Bojia Zi, Haibin Huang, Chi Zhang, and Xuelong Li. It is built on Qwen-Image-Edit-2509 and directly addresses the core challenge that style and content tend to entangle in Diffusion Transformer internal features. Training uses a curated-plus-wild triplet dataset scaled to tens of thousands of styles combined with a curriculum continual learning framework, and the model achieves state-of-the-art open-source results on style similarity, content consistency, and aesthetic quality.

Core capabilities come in two layers. First, image style transfer: feed one content image and one style reference, and the model produces a new image that preserves the original subject silhouette and composition in the chosen style. Second, video style transfer: the first frame is used as the content input, a style reference is fixed, and a v2v style transfer module keeps the frames temporally consistent — no jitter, no flicker. The repository ships telestyleimage_inference.py and telestylevideo_inference.py, plus a sample command — telestylevideo_inference.py --video_path ... --image_path ... --output_path results. Checkpoints are mirrored on Hugging Face at Tele-AI/TeleStyle; an online demo runs on the Space witcherderivia/TeleStyle; community ComfyUI nodes (aistudynow/Comfyui-tetestyle-image-video and neurodanzelus-cmd/ComfyUI-TeleStyle) plug TeleStyle into existing workflows.

Use cases include illustrators and graphic designers applying a consistent style to characters or product photos, short-video and self-media teams re-skinning their visual identity, comic and animation studios doing cross-frame style transfer, and game studios generating multiple skins for the same character. The technical report is published on arXiv as 2601.20175; TeleStyleV2 (self-distillation for artistic-content reference) was released in June 2026. Code and model weights are Apache-2.0. Deployment requires Python 3.11, PyTorch 2.9.1, CUDA 12.1, diffusers 0.36.0, and transformers 4.57.3. Note: GitHub reports the repository language as JavaScript because the GitHub Pages site inflates static-asset counts; the actual inference stack is PyTorch + diffusers (per the README install section).