Conditional Variational Diffusion Models

Gabriel della Maggiora,Luis Alberto Croquevielle,Nikita Deshpande,Harry Horsley,Thomas Heinis,Artur Yakimovich
2024-04-26
Abstract:Inverse problems aim to determine parameters from observations, a crucial task in engineering and science. Lately, generative models, especially diffusion models, have gained popularity in this area for their ability to produce realistic solutions and their good mathematical properties. Despite their success, an important drawback of diffusion models is their sensitivity to the choice of variance schedule, which controls the dynamics of the diffusion process. Fine-tuning this schedule for specific applications is crucial but time-costly and does not guarantee an optimal result. We propose a novel approach for learning the schedule as part of the training process. Our method supports probabilistic conditioning on data, provides high-quality solutions, and is flexible, proving able to adapt to different applications with minimum overhead. This approach is tested in two unrelated inverse problems: super-resolution microscopy and quantitative phase imaging, yielding comparable or superior results to previous methods and fine-tuned diffusion models. We conclude that fine-tuning the schedule by experimentation should be avoided because it can be learned during training in a stable way that yields better results.
Computer Vision and Pattern Recognition,Artificial Intelligence,Machine Learning
What problem does this paper attempt to address?
The paper proposes a Conditional Variational Diffusion Model (CVDM) for solving inverse problems, with a focus on the sensitivity of diffusion models to variance schedule selection when dealing with inverse problems. The current methods require manual adjustment of this schedule, while CVDM learns the schedule as part of the training process, allowing for data-driven conditioning and adaptation to different applications. By testing on two unrelated inverse problems (super-resolution microscopy and quantitative phase imaging), CVDM demonstrates results that are equal to or better than existing methods. The research found that schedule optimization can be learned during training, avoiding time-consuming manual tuning and providing improved solutions.