Latent Generative Models with Tunable Complexity for Compressed Sensing and other Inverse Problems

Researchers developed tunable-complexity priors for generative models (diffusion models, normalizing flows, and variational autoencoders) that can dynamically adjust complexity based on the specific inverse problem. The approach uses nested dropout and demonstrates superior performance across compressed sensing, inpainting, denoising, and phase retrieval tasks compared to fixed-complexity baselines.

• →Tunable-complexity priors consistently achieve lower reconstruction errors than fixed-complexity baselines across multiple inverse problem tasks.
• →The approach leverages nested dropout to dynamically adjust model complexity rather than using fixed dimensionality.
• →Theoretical analysis in linear denoising shows optimal tuning parameters depend on noise levels and model structure.
• →The method applies to multiple generative model types including diffusion models, normalizing flows, and variational autoencoders.
• →Research demonstrates potential for adaptive generative priors in solving inverse problems more effectively.