GCC: Generative Color Constancy via Diffusing a Color Checker

Extracting illumination's RGB color Input image Output image with color cast removed Image with an inpainted color checker Achromatic squares Color checker inpainting Inpainted color checker Color cast removed Inpainted color checker Color cast removed Inpainted color checker Color cast removed Angular error: 1.10 Angular error: 0.19 Angular error: 0.34 Figure 1. Our method performs color constancy through diffusion-based color checker inpainting. (top left) Given an input image, we first inpaint a color checker with Stable Diffusion, aligning the achromatic (gray) squares to accurately reflect the scene illumination (top middle). The RGB color extracted from the achromatic squares is then used to remove the color cast from the input image (top right). (Bottom) Our approach leverages the strong priors of pre-trained diffusion models to accurately estimate scene illumination without requiring physical color checkers during capture, enabling accurate white balance correction across diverse scenes.