Source-linked AI summary
TeleStyle: Content-Preserving Style Transfer in Images and Videos
Shiwen Zhang, Xiaoyan Yang, Bojia Zi, Haibin Huang, Chi Zhang, Xuelong Li
TL;DR
DiTs entangle content and style, and Qwen-Image-Edit lacks native dual-reference stylization. TeleStyle uses curriculum training on collected and synthesized triplets and a first-frame-conditioned video module; the report claims state-of-the-art performance in style similarity, content preservation, and aesthetic quality.
Problem
Diffusion Transformers entangle content and style features, while Qwen-Image-Edit lacks native support for dual-reference stylization.
Source: §1 · p. 1Method
TeleStyle trains with Curriculum Continual Learning on collected and synthesized triplets and uses a stylized first frame to propagate style across video sequences.
Source: §1 · p. 1, §3.1.1 · p. 4Results
TeleStyle achieves state-of-the-art performance in style similarity, content preservation, and aesthetic quality.
Source: §4.2 · p. 7, §5 · p. 8Takeaways & Limitations
The authors position TeleStyle as a robust baseline for future research in DiT-based style transfer and high-fidelity video editing.
Source: §5 · p. 8Training on collected triplets alone can miss subtle content details such as facial identity and struggle with out-of-distribution styles absent from that dataset.
Source: §3.1.2 · p. 5
Abstract
from arXiv · showhide
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
1 Introduction
Diffusion Transformers entangle content and style features, making content-preserving transfer challenging, and Qwen-Image-Edit lacks native dual-reference stylization. TeleStyle addresses this with hybrid-data curriculum training and a video module that propagates style from a stylized first frame.
1 Introduction
- DiT feature entanglement makes it non-trivial to preserve content structure while transferring style, and Qwen-Image-Edit lacks native support for dual-reference stylization.The report identifies Qwen-Image-Edit as a strong foundation for a data-centric approach.Source: §1 · p. 1
- Curriculum Continual Learning progresses from curated clean triplets to synthesized in-the-wild data to address limited generalization and noise-related content-fidelity trade-offs.The report says limited high-quality triplets restrict generalization, while large-scale in-the-wild data can degrade content fidelity.Source: §1 · p. 1
- A lightweight video module conditions on a stylized first frame and propagates style cues across the sequence to maintain temporal consistency.Figure 1 presents the image model accepting style and content references for content-preserving transfer while maintaining high aesthetic merit.Source: §1 · p. 1, Figure 1 · p. 2
2 Related Works
Prior image stylization work spans training-free UNet methods and conditional DiTs, while TeleStyle unifies content preservation and style transfer without style-specific LoRAs. Video approaches include key-frame propagation, paired-data training, and unified editing frameworks, with trade-offs in style alignment and fidelity.
2.1 Image Style Transfer
- Training-free UNet methods inject style or combine content and style encoders, while CSGO trains on content-style pairs.The cited approaches include StyleID, StyleShot, InstantStyle, and CSGO.Source: §2.1 · p. 2
- Conditional DiT methods support multiple reference or conditioning inputs, but Qwen Image Edit and FLUX-Kontext lack subject-plus-style references; TeleStyle unifies both capabilities without style-specific LoRAs.OmniStyle trains on style-reference, content-reference, and target triplets; OmniConsistency instead uses a separate consistency branch and external Style LoRAs.Source: §2.1 · p. 2
2.2 Video Stylization
- Video stylization spans training-free key-frame propagation, paired-data end-to-end models, and unified frameworks, whose generality can trade off against style fidelity and fine-grained texture control.Prior key-frame methods can inherit entangled style representations that misalign semantic structure and low-level texture, leading to artifacts or inconsistent stylization.Source: §2.2 · pp. 2–3
3 Methods
TeleStyle addresses scarce real-world style-transfer triplets with a hybrid dataset and staged training that balances content fidelity with generalization. Its video module propagates a reference image’s style across source frames using temporal conditioning.
3.1.1 Triplet Training Dataset Construction
- The hybrid dataset pairs 300k high-quality collected triplets with 1 million synthetic triplets generated by matching stylized images within the same style clusters, assigning one as reference and one as target.The collected dataset was resource-intensive to produce and contained only 30 distinct style categories.Source: §3.1.1 · p. 4
3.1.2 Content-Preserving Style Transfer via Curriculum Continual Learning
- Three-stage curriculum training moves from collected triplets, to a content-consistency-weighted subset, then to that subset mixed with a low ratio of synthetic data.The final stage is intended to improve out-of-distribution style generalization while maintaining acceptable content consistency; synthetic data alone can degrade content.Source: §3.1.2 · p. 5
3.1.3 Training and Inference Details
- Image training uses LoRA and rectified flow matching with style, content, and prompt conditioning; the default prompt is recommended for stable inference.The report also notes that matching the content-reference and output aspect ratios is crucial.Source: §3.1.3 · p. 5
3.2 Video Stylization
- For video stylization, a reference image and source video are encoded separately, then fused with noisy latents and processed by DiT blocks.The reference receives temporal index 0 as an anchor for style propagation; training uses flow matching on synthesized video stylization pairs and does not require optical-flow guidance or test-time optimization.Source: Figure 4 · p. 6, §3.2 · pp. 6–7
4 Experiments and Discussion
The experiments evaluate TeleStyle V1 across content, style, and aesthetic criteria, and report advantages over prior DiT-based style-transfer models. They also describe the training setup, benchmark, and metrics used for evaluation.
4.1 Implementation and Evaluation
- TeleStyle-Image is trained on Qwen-Image-Edit-2509 with LoRA rank 32, 1,024-pixel minimum edge, four H100 GPUs, batch size one per GPU, and a 1e-4 learning rate.Gradient checkpointing is used to save memory.Source: §4.1 · p. 7
- TeleStyle-Video combines Ditto data with an internal video collection and filters out clips with negligible motion to avoid temporally misaligned stylization supervision.The video model is trained on eight H100 GPUs with batch size four per GPU and a 1 × 10−5 learning rate.Source: §4.1 · p. 7
- The test benchmark pairs 50 style references with 40 content references to create 2,000 pairs, spanning diverse styles and complex content scenarios.The validation set contains 10 style references and 10 content references.Source: §4.1 · p. 7
- Style consistency uses CSD, aesthetics uses the LAION Aesthetics Predictor, and content preservation uses CPC, which zeros the CLIP content score when style consistency falls below a threshold.The content caption is generated by Qwen-VL, and its CLIP score against the output supplies the content score.Source: §4.1 · p. 7
4.2 Comparison with State-of-the-art Methods
- TeleStyle V1 shows superior style similarity and content consistency to prior models while maintaining high aesthetics, and the quantitative comparison reports significant advantages.The qualitative comparison is shown in Figure 5; no quantitative values are supplied here.Source: §4.2 · pp. 7–8, Figure 5 · p. 8
5 Conclusion
TeleStyle is presented as a framework for content-preserving style transfer in images and videos, with reported state-of-the-art performance across style similarity, content preservation, and aesthetic quality.
- TeleStyle is presented as a practical framework for content-preserving style transfer across images and videos, with state-of-the-art performance in style similarity, content preservation, and aesthetic quality.The report frames it as a potential baseline for DiT-based style transfer and high-fidelity video editing.Source: §5 · p. 8