Quantifying combanitorial capabilities of image-generating AI

• Main Author: Karanam, Akshit
• Other Authors: Li Boyang
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science
• Online Access: https://hdl.handle.net/10356/175738
• Institution: Nanyang Technological University

This study explores the combinatorial capabilities of Image-Generating AI technologies, specifically assessing the ability of these models to generate images based on the combination of multiple input objects. The objects used in the prompts were categorized based on their frequency of appearance in the training datasets, leading to the creation of various combination types. The primary aim was to determine which combination type leads to the most accurate image generation. To achieve this, language models were employed to construct prompts with selected objects, and the resulting images were generated and compared against actual photographs. The resemblance was quantified using the Fréchet Inception Distance (FID) scores.