From Missing Pieces to Masterpieces: Image Completion With Context-Adaptive Diffusion

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Image completion is difficult, with diffusion models struggling to blend generated content coherently with existing image parts.
• Existing methods lack explicit spatial and semantic alignment, leading to inconsistencies and poor integration.

Purpose of the Study:

• To introduce ConFill, a novel framework for high-fidelity image completion.
• To address the coherence and integration challenges in diffusion-based image completion.

Main Methods:

• Developed a Context-Adaptive Discrepancy (CAD) model to align intermediate distributions of known and unknown image regions.
• Implemented a Dynamic Sampling mechanism to adaptively refine sampling rates in complex regions.

Main Results:

• ConFill progressively reduces discrepancies between generated and original image content at each diffusion step.
• The Dynamic Sampling mechanism enhances detail and integration in restored areas.
• Extensive experiments show ConFill surpasses current state-of-the-art image completion methods.

Conclusions:

• ConFill establishes a new benchmark for image completion by achieving contextually aligned and seamlessly integrated results.
• The proposed CAD model and Dynamic Sampling mechanism effectively tackle coherence issues in generative models.