Uss, WojciechKaliński, WojciechKuznetsov, AlexandrAnand, HarishKim, SungyeCeylan, DuyguLi, Tzu-Mao2025-05-092025-05-092025978-3-03868-268-41017-4656https://doi.org/10.2312/egs.20251030https://diglib.eg.org/handle/10.2312/egs20251030Despite the latest advances in generative neural techniques for producing photorealistic images, they lack generation of multi-bounce, high-frequency lighting effect like caustics. In this work, we tackle the problem of generating cardioid-shaped reflective caustics using diffusion-based generative models. We approach this problem as conditional image generation using a diffusion-based model conditioned with multiple images of geometric, material and illumination information as well as light property. We introduce a framework to fine-tune a pre-trained diffusion model and present results with visually plausible caustics.Attribution 4.0 International LicenseCCS Concepts: Computing methodologies → Artificial intelligence; Neural networks; Image-based renderingComputing methodologies → Artificial intelligenceNeural networksImagebased rendering10.2312/egs.202510304 pages