Blind Inversion Using Latent Diffusion Priors.

Area of Science:

• Computational Imaging
• Machine Learning
• Computer Vision

Background:

• Diffusion models excel at modeling complex priors for inverse problems.
• Existing methods often require known forward operators, limiting practical use.
• Pixel-space diffusion models are common, but latent diffusion models (LDMs) offer greater potential.

Purpose of the Study:

• To introduce LatentDEM, a novel technique for blind inverse problems using latent diffusion priors.
• To address limitations of existing methods by handling unknown forward operators.
• To explore the application of LDMs in challenging inverse problems.

Main Methods:

• Developed LatentDEM, an iterative Expectation-Maximization (EM) framework for blind inverse problems.
• Incorporated LDM priors in the E-step for image recovery and forward operator estimation in the M-step.
• Proposed novel optimization techniques for LDM priors within the EM framework.

Main Results:

• LatentDEM effectively solves blind inverse problems by estimating unknown forward operators.
• Achieved accurate and efficient blind inversion results using tailored optimization techniques.
• Demonstrated superior performance on 2D blind deblurring and 3D pose-free sparse-view reconstruction tasks.

Conclusions:

• LatentDEM provides a general framework for linear and non-linear inverse problems.
• The method enables new capabilities in 3D inverse rendering.
• LatentDEM shows significant efficacy compared to prior art in complex imaging tasks.