Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TeleStyle: Content-Preserving Style Transfer in Images and Videos

Shiwen Zhang, Xiaoyan Yang, Bojia Zi, Haibin Huang, Chi Zhang, Xuelong Li

TL;DR

TeleStyle addresses the challenge of content-preserving style transfer within Diffusion Transformer architectures by leveraging the Qwen-Image-Edit platform and a Curriculum Continual Learning framework that combines a curated triplet dataset with large-scale synthetic data. The method introduces a three-stage training strategy to stabilize content fidelity while acquiring broad style generalization, and extends to video with a first-frame-conditioned propagation module to ensure temporal coherence. Empirical results demonstrate state-of-the-art performance in style similarity, content preservation, and aesthetics for both images and videos, supported by robust quantitative metrics and qualitative comparisons. The work advances practical high-fidelity style transfer suitable for real-world editing, with publicly available code and pretrained models.

Abstract

Content-preserving style transfer, generating stylized outputs based on content and style references, remains a significant challenge for Diffusion Transformers (DiTs) due to the inherent entanglement of content and style features in their internal representations. In this technical report, we present TeleStyle, a lightweight yet effective model for both image and video stylization. Built upon Qwen-Image-Edit, TeleStyle leverages the base model's robust capabilities in content preservation and style customization. To facilitate effective training, we curated a high-quality dataset of distinct specific styles and further synthesized triplets using thousands of diverse, in-the-wild style categories. We introduce a Curriculum Continual Learning framework to train TeleStyle on this hybrid dataset of clean (curated) and noisy (synthetic) triplets. This approach enables the model to generalize to unseen styles without compromising precise content fidelity. Additionally, we introduce a video-to-video stylization module to enhance temporal consistency and visual quality. TeleStyle achieves state-of-the-art performance across three core evaluation metrics: style similarity, content consistency, and aesthetic quality. Code and pre-trained models are available at https://github.com/Tele-AI/TeleStyle

TeleStyle: Content-Preserving Style Transfer in Images and Videos

TL;DR

TeleStyle addresses the challenge of content-preserving style transfer within Diffusion Transformer architectures by leveraging the Qwen-Image-Edit platform and a Curriculum Continual Learning framework that combines a curated triplet dataset with large-scale synthetic data. The method introduces a three-stage training strategy to stabilize content fidelity while acquiring broad style generalization, and extends to video with a first-frame-conditioned propagation module to ensure temporal coherence. Empirical results demonstrate state-of-the-art performance in style similarity, content preservation, and aesthetics for both images and videos, supported by robust quantitative metrics and qualitative comparisons. The work advances practical high-fidelity style transfer suitable for real-world editing, with publicly available code and pretrained models.

Abstract

Content-preserving style transfer, generating stylized outputs based on content and style references, remains a significant challenge for Diffusion Transformers (DiTs) due to the inherent entanglement of content and style features in their internal representations. In this technical report, we present TeleStyle, a lightweight yet effective model for both image and video stylization. Built upon Qwen-Image-Edit, TeleStyle leverages the base model's robust capabilities in content preservation and style customization. To facilitate effective training, we curated a high-quality dataset of distinct specific styles and further synthesized triplets using thousands of diverse, in-the-wild style categories. We introduce a Curriculum Continual Learning framework to train TeleStyle on this hybrid dataset of clean (curated) and noisy (synthetic) triplets. This approach enables the model to generalize to unseen styles without compromising precise content fidelity. Additionally, we introduce a video-to-video stylization module to enhance temporal consistency and visual quality. TeleStyle achieves state-of-the-art performance across three core evaluation metrics: style similarity, content consistency, and aesthetic quality. Code and pre-trained models are available at https://github.com/Tele-AI/TeleStyle
Paper Structure (24 sections, 4 equations, 5 figures, 1 table, 1 algorithm)

This paper contains 24 sections, 4 equations, 5 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. Image Style Transfer
  4. Zero-Shot Style Transfer with UNet Diffusion Models
  5. Zero-Shot Style Transfer with Conditional DiT
  6. Video Stylization
  7. Methods
  8. Content-Preserving Style Transfer on Images
  9. Triplet Training Dataset Construction
  10. Content-Preserving Style Transfer via Curriculum Continual Learning
  11. Stage 1: Capability Activation ($D_1$).
  12. Stage 2: Content Fidelity Refinement ($D_2$).
  13. Stage 3: Robust Generalization ($D_3$).
  14. Training and Inference Details
  15. Video Stylization
  16. ...and 9 more sections

Figures (5)

  • Figure 1: TeleStyle-Image accepts style and content references for content-preserving style transfer, while maintaining high aesthetics merit. TeleStyle-Image is the first content-preserving style transfer model built on Qwen-Image-Edit.
  • Figure 2: TeleStyle-Video takes a stylized first frame and a source video as input and propagates the style coherently across the entire sequence. It achieves high-quality results across a diverse range of artistic styles—including oil painting, sci-fi, and watercolor—while preserving temporal consistency. Notably, it remains effective even under challenging conditions such as anime-style rendering, where large structural and contour variations are present.
  • Figure 3: Overview of dataset construction: Collected triplets $D_{collected}$ (left) and synthetic triplets $D_{synthetic}$ (right).
  • Figure 4: Overview of the video stylization architecture. A reference image and source video frames are encoded via dual Patch Embedders, fused with noisy latents, and processed by DiT blocks alongside empty text tokens.
  • Figure 5: Qualitative Comparison with State-of-the-art Style Transfer Models.