Deep data consistency: A fast and robust diffusion model-based solver for inverse problems.

Area of Science:

• Computer Vision
• Machine Learning
• Image Reconstruction

Background:

• Diffusion models offer a powerful prior for image inverse problems.
• Existing methods struggle to balance data consistency and image realness, with slow sampling speeds being a key limitation.

Purpose of the Study:

• To introduce Deep Data Consistency (DDC), a novel approach to enhance diffusion models for image inverse problems.
• To address the challenges of data consistency, image realness, and sampling speed in diffusion model applications.

Main Methods:

• DDC integrates a deep learning model into the data consistency step of diffusion models.
• A variational bound training objective is employed to maximize the conditional posterior and minimize its impact on the diffusion process.

Main Results:

• DDC achieves superior performance in similarity and realness metrics compared to state-of-the-art methods on linear and non-linear tasks.
• High-quality solutions are generated rapidly, with an average of only 5 inference steps taking 0.77 seconds.

Conclusions:

• DDC demonstrates robust performance across various datasets, noise levels, and multiple tasks using a single pre-trained model.
• The proposed method significantly improves the efficiency and effectiveness of diffusion models for image inverse problems.