Virtual Consistency Model for All-in-one Image Restoration.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• All-in-one Image Restoration (AIR) aims for unified models to handle diverse image degradations.
• Existing methods face gradient conflicts due to degradation-specific training.
• Diffusion models offer a promising approach by operating in a high-noise space, homogenizing degradations.

Purpose of the Study:

• To address the fidelity compromisation in diffusion-based AIR methods caused by lack of direct image-space supervision.
• To resolve the dilemma where the optimal space for modeling degradations is suboptimal for image fidelity.
• To propose a novel method for effective All-in-one Image Restoration.

Main Methods:

• Introduced the Virtual Consistency Model for AIR (VCMAIR).
• VCMAIR restores images in a high-noise space.
• Employs a novel consistency function for accurate image-space supervision.

Main Results:

• VCMAIR outperforms existing state-of-the-art methods.
• Demonstrated superior performance across a comprehensive benchmark of diverse degradation scenarios.
• Successfully applied to both standard AIR tasks and challenging real-world image restoration.

Conclusions:

• The proposed VCMAIR effectively resolves the fidelity-supervision dilemma in diffusion-based AIR.
• VCMAIR offers a robust solution for diverse and challenging image restoration tasks.
• The novel consistency function enables accurate supervision in the image space while restoring in the noise space.