Grant: Discovery Launch Supplement

Text-to-image synthesis technology aims to convert textual descriptions into corresponding visual representations or images using advanced artificial intelligence models. The evolution of text-to-image synthesis has been marked by significant advancements, which have opened up exciting possibilities in various fields, including healthcare, education, digital art, and beyond. However, despite these advances, a significant challenge persists - the assessment of the quality and realism of AI-generated images and their alignment with textual descriptions. Evaluating the output of text-to-image synthesis models is crucial for ensuring their utility and reliability. Unfortunately, existing evaluation metrics are often imprecise and lack human-centricity, making it challenging to validate these models effectively. The absence of robust evaluation methods has been a significant obstacle to deploying these models in real-world applications, thus limiting their impact. In response to this challenge, the current proposal seeks to address three interconnected objectives: 1) Photorealism Assessment Metrics: Develop computational and subjective metrics that enable accurate assessment of the photorealism of AI-generated images. This will involve creating computational tools and methodologies that can measure the extent to which these images resemble real-world objects or scenes. 2) Comprehensive Image Quality Evaluation: Develop comprehensive evaluation tools that consider a range of image quality factors, allowing for a more holistic assessment of the quality of AI-generated images. 3) Text-to-Image Alignment Measures: Establish standardized methods for measuring the alignment between textual descriptions and the images they generate. This is crucial for ensuring that the visual output faithfully represents the intended content. The subjective tools developed in this program will undergo extensive reliability and validity testing. Furthermore, computational metrics developed within this program will be validated through comparisons with human evaluations and benchmarking against other computational metrics. By achieving these objectives, the current research aims to provide researchers, developers, and industries with the essential tools required to advance and standardize text-to-image synthesis technology. These tools will not only enhance the quality and reliability of AI-generated images but will also accelerate the adoption of this technology across various sectors.