Generative diffusion models have recently revolutionized computer vision due to their ability to create realistic, detailed images. By using billion-size training datasets these models learn to map noise to images based on text prompts. Multiple recent works reported that the opposite map (from signal to noise) captures rich information about images, allowing for applications like image editing, out-of-distribution correction, and 3D shape generation. In practice, however, these inversion maps are shown to have weak numerical stability. Our research aims to analyze the numerical errors of diffusion inversion in multiple practical settings. If successful, this has the potential to greatly improve multiple diffusion-based methods and enhance understanding of generative diffusion models.

I am participating in SuperUROP because I’m eager to apply what I’ve learned from machine learning (ML) courses to an extended research project. As ML techniques have advanced, I’ve become deeply interested in being able to explain why these methods work intuitively from a mathematical standpoint, and I’m excited that my project will both let me explore my interest while gaining experience in conducting research especially in computer vision.