Personalized Image Filter:
Mastering Your Photographic Style

Chengxuan Zhu¹, Shuchen Weng², Jiacong Fang³, Peixuan Zhang⁴,
Si Li⁴, Chao Xu¹, Boxin Shi^{3, *}

¹State Key Laboratory of General Artificial Intelligence, School of Intelligence Science and Technology, Peking University ²Beijing Academy of Artificial Intelligence ³ State Key Laboratory of Multimedia Information Processing and National Engineering Research Center of Visual Technology, School of Computer Science, Peking University ⁴ School of Artificial Intelligence, Beijing University of Posts and Telecommunications ^*Corresponding author

arXiv Code (Coming Soon)

Comparison Results

* The text prompt is only used for text-based image editing. Proposed method (PIF) does not use the text prompts as input.

Left Method

Right Method

Reference Style

Loading prompt...

Text Prompt*

Left Method

Right Method

Reference Style

Loading prompt...

Text Prompt*

Left Method

Right Method

Reference Style

Loading prompt...

Text Prompt*

Results with Residual One-step Diffusion

After finetuning with the proposed residual one-step diffusion paradigm, the model demonstrates the ability to adjust the photographic style conditioned on the given text prompt, while maintaining the fidelity of the original image content.

Content

Comparison with Average

Loading prompt...

Average Style Prompt (Left)

Loading prompt...

Modified Style Prompt (Right)

Incremental Results

With the concepts optimized from reference image, PIF renders them onto content image incrementally.

Content

+Tint

Incremental Result

Reference

+Tint +Sharpness +Highlight +Saturation +Shadow +Contrast +Exposure +Vignet.

Results with Conflicting Concepts

Reference images may contain conflicting concepts beyond the shared style. PIF learns them like vector summation: since optimization only updates pseudo-word embeddings, and word embeddings are roughly linear, conflicting concepts cancel out, leaving embeddings close to an “average.”

Content

Reference A

Result

Reference A†

Reference B†

Reference A Reference A+B^# Reference B

†: The image of Reference A shows a blue tint, light shadow, low contrast, and fuzzy highlight. Reference B shares the photographic style in contrast, highlight, sharpness and shadow, but differs in the tint, by leaning towards orange.

^#: Reference A+B means the text embeddings are optimaized on both Reference A and Reference B

BibTeX

@article{zhu2025pif,
      title={Personalized Image Filter: Mastering Your Photographic Style}, 
      author={Chengxuan Zhu and Shuchen Weng and Jiacong Fang and Peixuan Zhang and Si Li and Chao Xu and Boxin Shi},
      journal={arXiv preprint arXiv: 2510.16791},
      url={https://arxiv.org/abs/2510.16791}, 
      year={2025},
}