Automated Quality Assessment for AIGC Image Generation: Recent Research Advances

Quality assessment is crucial for AIGC image generation to ensure content professionalism, commercial value, and customer satisfaction. Traditional manual evaluation is time-consuming, labor-intensive, and subject to subjective bias, making automated quality assessment methods an urgent need. This article shares several latest research advances in this field.

The first project focuses on aesthetic quality assessment of artistic images. It proposes a clear framework to quantify aesthetic scores in art images, constructs a multi-attribute, multi-category painting dataset called APDD (Art Paintings Dataset), and introduces a painting image evaluation network (AANSPS). The model achieves satisfactory results across most metrics, validating the effectiveness of the approach.

The second project addresses the challenge of evaluating generated images using existing metrics like Inception Score (IS) and Fréchet Inception Distance (FID), which don't well reflect human preferences. It builds a large-scale Human Preference Dataset v2 (HPD v2) with human-annotated preferences and trains a benchmark model Human Preference Score v2 (HPS v2) to measure generative algorithm development. HPS v2 performs better than previous metrics across various image distributions and can be used to improve text-to-image models.

The third project creates Pick-a-Pic, a web application that allows users to generate images and specify their preferences. This resulted in a large open dataset with user preferences, which was used to train PickScore, a CLIP-based scoring function that demonstrates excellent performance in predicting human preferences. PickScore can rank multiple generated images to compare different text-to-image models or the same model under different parameters.

These automated quality assessment approaches can quickly score images, save labor costs, and more accurately compare model performance to facilitate model iteration. The team is currently working on AIGC research in the home improvement industry to enhance AI model effectiveness and welcomes collaboration with interested researchers.