Speaker
Description
Given the challenges in LArTPC event reconstruction, we present the first steps toward a fully automated inference pipeline mapping 2D detector images to event properties. Inspired by the success of denoising diffusion probabilistic models (DDPMs) in natural image generation, we developed a modified latent diffusion model capable of conditionally generating single-particle LArTPC images with quality comparable to Geant4 simulations. Utilizing this model, we constructed an iterative matching process in which observed detector images are meaningfully compared with generated images conditioned on known particle characteristics (PID & momentum), enabling a data-driven approach to inference of event properties without the use of traditional reconstruction techniques.