APDDv2-SynAE

Project Summary

The APDDv2-SynAE (Aesthetics of Paintings and Drawings Dataset – Synthetic Aesthetic Evaluation) is an extension of the APDDv2, proposed by (Jin et al., 2024), developed to investigate and automate the aesthetic analysis of AI-generated images. For its construction, we performed a sampling of approximately 500 images from APDDv2, preserving the statistical relevance of the set. Each image was described using DeepSeek Janus, and, from these textual descriptions, we generated synthetic images in the same style and content as the originals using two models of different sizes:

• SMALL - Deepseek Janus Pro 1B
• BIG - Deepseek Janus Pro 7B

The goal was to reproduce the works from the original dataset solely from their descriptions, based on the premise that the generated images should present an aesthetic score similar to that of the original work. The evaluation was conducted using ArtCLIP, a model trained on the original dataset and proposed by the APDDv2 authors, enabling the comparison of aesthetic criteria between human and synthetic images. Presented at ICCC (International Conference on Computational Creativity) 2025 through the paper: Automatic Aesthetic Evaluation in Generative Image Models, the APDDv2-SynAE expands the possibilities for comparative studies between human-created and AI-generated works, contributing to advancements in computational art and computer vision.

Methodology pipeline

Presentation video

Power Point Presentation

The file used for the ICCC presentation can be accessed at the link below.

Access Presentation

Dataset Viewer

Browse database samples through an interactive visual carousel.

Acess Viewer

Source Code and Repository

To access the source code used to create the dataset and other experiments in the project, access the repository on GitHub.

See on GitHub

Na Mídia

• [Estética de imagens por IA: aluna do DCC/UFMG apresenta estudo em conferência internacional]