Learning Personalized Photographic Style
from Pairwise User Preferences

Jinwoo Kim1 Jihye Yoo2,* Seon Joo Kim1
1CIPLAB, Yonsei University 2Ewha Womans University
*Work done during an internship at CIPLAB
Paper Dataset
Pairwise comparison interface: two photos of the same scene with different styles, asking the user to pick the preferred one.
PPS learning from pairwise preferences. Users repeatedly choose between two differently-styled renderings of the same scene. Each choice reveals a direction in color and tonal preference, and the model learns to generalize that direction to new images.

Abstract

Photographic style preferences are deeply personal, varying across individuals in color and tonal aesthetics. We introduce Personalized Photographic Style (PPS) learning, where the goal is to capture a user's implicit preferences from comparative judgments and apply them consistently across diverse images. To establish a foundation for this problem, we present three contributions. First, we introduce PPSD, a dataset containing pairwise preference judgments from 767 users, each providing an average of 70 comparisons. To capture diverse style signals, images are sourced from professional edits, device pipelines, and generative models. Second, we explore several baseline models demonstrating the feasibility of adapting style transfer and enhancement approaches for preference learning. Third, we develop a comparative evaluation framework suited to the implicit nature of personal preferences.

PPSD Dataset

A large-scale collection of pairwise user preferences for photographic style. 767 users, ~60,000 valid preference judgments, 1,192 unique scenes, and 7,972 unique image style pairs across five source categories.

Sample image pairs from PPSD across five source categories.
Five source categories. Professional expert edits, device processing variations, generative model outputs, synthetic adjustments, and content-based variations — together spanning a wide range of color and tonal style signals.
Dataset statistics: consistency, age, region, device, eyewear distributions.
Participant statistics. Consistency scores, age, geographic regions, display devices, and eyewear usage across 767 retained users.

Download the dataset

PPSD is hosted on Google Drive. Includes image pairs and user preference annotations.

Download

Baseline Methods

We explore three representative approaches for adapting existing models to PPS learning: a User-specific Decoder, a User Preference Embedding, and an Exemplar-based Inference model.

Overview of the inference pipeline and three baseline methods.
Inference pipeline and three baseline methods. (a) User-specific Decoder — one decoder per user. (b) User Preference Embedding — a global decoder conditioned on per-user embeddings. (c) Exemplar-based Inference — a transformer that conditions on preference triplets.

Qualitative Results

Qualitative results of baseline methods on PPSD.
Given either non-preferred or preferred images as input, models adjust toward the user's preferred aesthetic while preserving the preferred ones. The bottom row shows a cross-model comparison on identical scenes.

More in the Paper

This page is a quick tour. The paper goes deeper on dataset construction, modeling, evaluation, and what we learned along the way.

Read the full paper →

BibTeX

@inproceedings{kim2026pps,
  title     = {Learning Personalized Photographic Style from Pairwise User Preferences},
  author    = {Kim, Jinwoo and Yoo, Jihye and Kim, Seon Joo},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision
               and Pattern Recognition (CVPR)},
  year      = {2026},
}