PPR10K: A Large-Scale Portrait Photo Retouching Dataset With Human-Region Mask and Group-Level Consistency

Jie Liang, Hui Zeng, Miaomiao Cui, Xuansong Xie, Lei Zhang

Introduction

Professional portrait photography involves a sophisticated retouching process to transform raw sensor data into aesthetically pleasing images. While automated photo enhancement has seen significant progress, portrait-specific retouching remains challenging due to two primary factors: the need for human-region priority (HRP) and group-level consistency (GLC). Standard models often fail to prioritize the human subject over the background or maintain a unified aesthetic across a series of photos from the same session. To address these issues, researchers have introduced PPR10K, a large-scale dataset specifically curated for portrait retouching.

The PPR10K Dataset

The PPR10K dataset consists of 11,161 high-quality raw photos organized into 1,681 groups. Unlike general-purpose datasets like MIT-Adobe FiveK, PPR10K is tailored for portraits, featuring a diverse range of subjects, lighting conditions, and camera brands. Each image is accompanied by high-resolution human-region segmentation masks and has been professionally retouched by three different experts, providing multiple ground-truth targets for learning various aesthetic styles.

Example of a portrait photo from the PPR10K dataset with its corresponding human-region mask.

Methodology

To leverage the PPR10K dataset effectively, the researchers proposed two key learning strategies designed to mimic professional workflows:

Human-Region Priority (HRP)

In portraiture, the subject's appearance is paramount. The team introduced human-centered measures that utilize the provided segmentation masks to weight loss functions (e.g., Mean Squared Error) more heavily on pixels within the human region. This ensures that the model prioritizes skin tones and facial details, even if it requires a trade-off in background fidelity.

Group-Level Consistency (GLC)

Maintaining a consistent "look" across a photo set is crucial for professional albums. The framework implements a GLC learning strategy using a data augmentation technique. By taking overlapping crops of a single image and applying different random tonal adjustments, the model is trained to produce identical outputs for the overlapping regions. This approach teaches the model to ignore incidental lighting variations and maintain a stable aesthetic across varied frames.

A full-body portrait and its human-region mask, illustrating the dataset's focus on subject-specific retouching.

Results and Discussion

Experiments demonstrate that models trained on PPR10K significantly outperform those trained on general datasets for portrait tasks. Quantitatively, the HRP loss leads to improved perceptual quality on human subjects, while the GLC strategy effectively reduces color variance across photo groups. Furthermore, by employing 3D Look-Up Tables (3D LUTs), the framework achieves high-speed processing (over 500 fps), making it suitable for real-time production environments.

Conclusion

PPR10K provides a robust foundation for advancing automated portrait retouching. By emphasizing human-region priority and group-level consistency, this work bridges the gap between general photo enhancement and the specialized needs of professional portrait photography. Future research will explore more granular segmentation and local editing techniques to further refine the retouching process.

Summary updated

Abstract

Different from general photo retouching tasks, portrait photo retouching (PPR), which aims to enhance the visual quality of a collection of flat-looking portrait photos, has its special and practical requirements such as human-region priority (HRP) and group-level consistency (GLC). HRP requires that more attention should be paid to human regions, while GLC requires that a group of portrait photos should be retouched to a consistent tone. Models trained on existing general photo retouching datasets, however, can hardly meet these requirements of PPR. To facilitate the research on this high-frequency task, we construct a large-scale PPR dataset, namely PPR10K, which is the first of its kind to our best knowledge. PPR10K contains $1, 681$ groups and $11, 161$ high-quality raw portrait photos in total. High-resolution segmentation masks of human regions are provided. Each raw photo is retouched by three experts, while they elaborately adjust each group of photos to have consistent tones. We define a set of objective measures to evaluate the performance of PPR and propose strategies to learn PPR models with good HRP and GLC performance. The constructed PPR10K dataset provides a good benchmark for studying automatic PPR methods, and experiments demonstrate that the proposed learning strategies are effective to improve the retouching performance. Datasets and codes are available: https://github.com/csjliang/PPR10K.
BibTeX
@inproceedings{liang2021ppr,
  author = {Jie Liang and Hui Zeng and Miaomiao Cui and Xuansong Xie and Lei Zhang},
  title = {{PPR10K: A Large-Scale Portrait Photo Retouching Dataset With Human-Region Mask and Group-Level Consistency}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2021},
  doi = {10.1109/cvpr46437.2021.00071},
}
Figure