Personalized Reward Modeling for Text-to-Image Generation

Yonsei University
ECCV 2026
Motivation for personalized text-to-image evaluation with user-specific criteria.

Universal reward models often miss the visual criteria that matter to individual users. PIGReward learns those criteria from a few pairwise choices and uses them to compare new text-to-image outputs.

Abstract

Recent text-to-image models generate semantically coherent images from textual prompts, yet evaluating how well they align with individual user preferences remains an open challenge. Conventional evaluation methods, including general reward functions and similarity-based metrics, fail to capture the diversity and complexity of personal visual tastes.

We present PIGReward, a personalized reward modeling framework that dynamically generates user-conditioned evaluation dimensions and assesses images through structured chain-of-thought reasoning. PIGReward uses self-bootstrapping over limited reference data to construct rich user contexts, enabling personalization without user-specific training. Beyond evaluation, PIGReward provides personalized feedback for prompt optimization, improving alignment between generated images and individual intent.

Method

PIGReward decomposes personalized evaluation into two learned components. A preference reasoner π converts each user reference pair into an explicit natural-language rationale. A reward model φ then induces user-specific evaluation dimensions from those rationales and performs dimension-wise comparison for a target image pair.

Bootstrapping Preference Rationales

The preference reasoner learns to produce faithful rationales by contrasting correct and incorrect explanations for the same image pair.

Training process for the preference reasoner with contrastive DPO rationales.

Training the Reward Model

The reward model is distilled from filtered chain-of-thought evaluation trajectories, learning to score image pairs with user-conditioned dimensions.

Training process for the personalized reward model with CoT evaluation data and supervised fine-tuning.

PIGBench

Examples from PIGBench with user-specific rankings.

We introduce PIGBench, a user-study benchmark for realistic personalized text-to-image evaluation. Each instance contains a user-specific ranking of four SDXL-generated images from one abstract prompt.

PIGBench is built from 100 curated image sets and 75 user records, with reference lengths varying from 5 to 15 examples to test personalization under sparse feedback.

Results

PIGReward consistently outperforms conventional reward models, LVLM-based judges, and reference-conditioned baselines across personalized T2I evaluation benchmarks.

Method Pick-a-Pic PIP PASTA PIGBench
UnifiedReward-Think 59.92 35.28 65.71 66.04
Qwen2.5-VL-7B w/ reference 61.52 49.95 50.34 54.72
GPT-4o w/ reference 46.90 76.23 68.80 64.15
PIGReward 63.76 77.84 75.43 84.91

Accuracy with tie is shown for each benchmark.

Why Similarity Is Not Enough

Similarity metrics can overfit to surface resemblance between the target image and prior liked images. PIGReward instead infers the user's underlying evaluative criteria, such as detail, composition, symbolism, or realism, and applies them to the target pair.

Failure case of similarity-based metrics for personalized evaluation.

Personalized Prompt Optimization

Overview of personalized prompt optimization using PIGReward.
Win rates for personalized prompt optimization.

When used as a user-conditioned reward function, PIGReward provides preference feedback for prompt model training. The optimized prompts produce images that align more closely with individual visual preferences than generic prompt optimization.

Qualitative Study

Qualitative results for personalized prompt optimization across users and baselines.

BibTeX

@article{lee2025personalized,
  title={Personalized Reward Modeling for Text-to-Image Generation},
  author={Lee, Jeongeun and Heo, Ryang and Lee, Dongha},
  journal={arXiv preprint arXiv:2511.19458},
  year={2025}
}