Personalized Reward Modeling for Text-to-Image Generation (ECCV 2026)

PIGReward is a personalized reward model for text-to-image generation. Given a text prompt, two generated images, and a user's preference context, it compares the image pair using user-specific evaluation criteria and predicts which image better matches the user's visual taste.

This repository contains the Context-to-Reward evaluator.

For bootstrapping preference context from prior image choices, use:

https://huggingface.co/jeongeunnn/pigreward-bootstrap

Model

Base model: Qwen2-VL-7B-Instruct
Task: Personalized pairwise image preference evaluation
Input: Text prompt, two candidate images, and user preference context
Output: Dimension-wise reasoning, scores, and final preferred image

Usage

import torch
from transformers import AutoProcessor, AutoTokenizer, Qwen2VLForConditionalGeneration

model_id = "jeongeunnn/pigreward"

processor = AutoProcessor.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=True)
model = Qwen2VLForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model.eval()

Example Output

Composition: Image 1 shows stronger spatial balance and a clearer focal structure. (Score 1: 8, Score 2: 6)
Lighting: Image 2 has warmer and more atmospheric lighting. (Score 1: 6, Score 2: 8)
Detail: Image 1 contains sharper textures and more coherent visual details. (Score 1: 9, Score 2: 7)
Score 1: 23
Score 2: 21
Image 1 is better

Citation

@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}
}