Enlightening Photographic Style Transfer with a Self-Supervised Photographic Embedding
Chengxuan Zhu* · Jiacong Fang* · Shuchen Weng · Youwei Lyu · Jiajun Tang · Qingnan Fan · Chao Xu · Boxin Shi
*: Equal Contribution
Photographic style — the nuanced play of lightness, color, and tone a photographer crafts — is easy for the eye to read, yet invisible to most image embeddings. We present PETAL (Photographic Embedding for Transfer with an Adaptive LUT): we learn a continuous photographic embedding by self-supervision, and use it to drive a lightweight adaptive neural LUT that transfers style faithfully, with no test-time optimization.
Quick Start
1. Environment Setup
# A CUDA GPU is strongly recommended; CPU inference also works but is much slower.
# Install uv (if not already installed)
curl -LsSf https://astral.sh/uv/install.sh | sh
uv sync # Create the virtual environment and install dependencies
2. Download Checkpoints
The released weights are ckpts/style_encoder.pt and ckpts/lut_model.pt in
this repository. To fetch them into a local ./ckpts folder, clone the
code repository and run its download
helper:
./download_weights.sh
You can also download the two files from the Files tab and place them in
./ckpts.
3. Inference
Content images and style images are paired by file stem (filename without
extension): content_dir/001.jpg is paired with style_dir/001.png.
uv run python infer.py \
--content_dir ./data/test_pairs/content \
--style_dir ./data/test_pairs/reference \
--output_dir ./outputs \
--output_format png
| Argument | Default | Description |
|---|---|---|
--content_dir | (required) | Directory of content images |
--style_dir | (required) | Directory of style images |
--output_dir | ./outputs | Output directory |
--lut_model_path | ./ckpts/lut_model.pt | Path to LUT / AdaIN checkpoint |
--style_model_path | ./ckpts/style_encoder.pt | Path to style encoder checkpoint |
--style_input_size | 512 | Resolution for the style encoder |
--guidance_scale | 1.0 | Guidance scale for adain, suggested range [0.5, 1.2] |
--mixed_precision | no | no / fp16 / bf16 |
--output_format | png | png / jpg / jpeg (PNG is lossless; JPG/JPEG are lossy) |
--skip_existing | off | Skip pairs whose output already exists |
--seed | 42 | Random seed |
Citation
@inproceedings{zhu2026photographic,
title = {Enlightening Photographic Style Transfer with a Self-Supervised Photographic Embedding},
author = {Zhu, Chengxuan and Fang, Jiacong and Weng, Shuchen and Lyu, Youwei and Tang, Jiajun and Fan, Qingnan and Xu, Chao and Shi, Boxin},
booktitle = {Proceedings of the European Conference on Computer Vision},
year = {2026}
}