DiffGIC: Diffusion Prior Based Null-Space Correction for High Resolution Grayscale Image Colorization

Diffusion models have demonstrated exceptional abilities in colorizing grayscale images. To colorize high-resolution images, current methods use a strategy that combines super-resolution with hierarchical image processing (SR-HIPS). This approach involves shrinking the input images for the diffusion model to reduce computational resources. However, this can lead to the loss of detailed information in high-resolution grayscale images. To overcome this limitation, this paper introduces DiffGIC, a novel method leveraging color image decomposition via range-null space decomposition. DiffGIC takes color from low-resolution color images and details from high-resolution grayscale images. By adjusting the color using a pre-trained diffusion model and combining it with detailed grayscale information, our method produces high-quality, high-resolution colored images. Moreover, DiffGIC improves upon the SR-HIPS strategies by adding detailed grayscale details into the colorization process, marking a notable advancement over previous methods.