Fig 1.
Results generated by the SDXL with our approach UltraStyle.
Reprinted from [http://xhslink.com/o/nfgAm1FZ5w] under a CC BY license, with permission from Xiaoming Huang, original copyright 2024. Reprinted from [https://pan.baidu.com/s/10_CjlBAaXZ6vB_RONzfU3g?pwd=b97i] under a CC BY license, with permission from Yi Yang, original copyright 2025.
Fig 2.
To facilitate the training of both style and content LoRA modules, we replace the conventional prediction paradigm with a novel prediction formulation. For content LoRA training, we design a loss transition mechanism that simultaneously captures the global structural layout and the fine-grained local details of the content image. To effectively disentangle the style and content information encoded in the style image, we adopt a two-stage training framework: first, we optimize a content-consistent LoRA module via the proposed loss transition; subsequently, we freeze the content LoRA and train a dedicated style LoRA to encode stylistic variations independently. Reprinted from [https://pan.baidu.com/s/10_CjlBAaXZ6vB_RONzfU3g?pwd=b97i] under a CC BY license, with permission from Yi Yang, original copyright 2025.
Fig 3.
We present style transfer results of our method and three baseline methods.
Reprinted from [https://pan.baidu.com/s/10_CjlBAaXZ6vB_RONzfU3g?pwd=b97i] under a CC BY license, with permission from Yi Yang, original copyright 2025. Reprinted from [http://xhslink.com/o/9GlJdtObddi] under a CC BY license, with permission from Wuwei Zhang, original copyright 2025. Reprinted from [http://xhslink.com/o/54Eeo1PlMG] under a CC BY license, with permission from Tao Pu, original copyright 2025.
Table 1.
Quantitative comparison of style and content alignment. Lower DS and higher CLIP/DINO indicate better performance.
Fig 4.
Visualization Results of the Ablation Study.
Reprinted from [https://pan.baidu.com/s/10_CjlBAaXZ6vB_RONzfU3g?pwd=b97i] under a CC BY license, with permission from Yi Yang, original copyright 2025. Reprinted from [http://xhslink.com/o/9GlJdtObddi] under a CC BY license, with permission from Wuwei Zhang, original copyright 2025.
Table 2.
Ablation study on the dual-phase training strategy.
Table 3.
Ablation study on the progressive loss transition strategy.
Table 4.
Ablation study on the Decoupled Inference Controller.
Fig 5.
Visualization Results of Different .
Reprinted from [https://pan.baidu.com/s/10_CjlBAaXZ6vB_RONzfU3g?pwd=b97i] under a CC BY license, with permission from Yi Yang, original copyright 2025.