Blockwise Divide-and-Aggregate for Image Restoration using Diffusion Priors

Diffusion models have emerged as powerful generative priors for solving linear inverse problems in image restoration. However, directly applying these models to high-resolution inputs is challenging due to their fixed training resolution and the computational cost of full-image denoising trajectories. In this work, we propose a Blockwise Divide-and-Aggregate framework that decomposes the image restoration task into spatially localized subproblems, each solved using independent diffusion model trajectories, and then aggregates the block-level solutions into a globally consistent reconstruction. By running multiple generative trajectories in parallel over image blocks, our approach efficiently leverages pre-trained diffusion priors without retraining, while naturally scaling to high-resolution images. We demonstrate that the proposed aggregation strategy effectively handles overlapping block boundaries and preserves global consistency across the restored image. Experiments on standard image restoration benchmarks demonstrate state-of-the-art performance across a range of linear inverse problems, including inpainting, super-resolution, and deblurring.