GCC: Generative Color Constancy via Diffusing a Color Checker
Chen-Wei Chang, Cheng-De Fan, Chia-Che Chang, Yi-Chen Lo, Yu-Chee Tseng, Jiun-Long Huang, Yu-Lun Liu
2025-02-25
Summary
This paper talks about GCC, a new method for making colors in photos look right no matter what camera took them, by using AI to add a special color checker into the image.
What's the problem?
Different cameras see colors differently, which makes it hard for computers to figure out the right colors in photos. This is especially tricky when you want a system that works well with many different types of cameras.
What's the solution?
The researchers created GCC, which uses AI to imagine and add a color checker (like the ones photographers use) into any photo. This checker helps the system figure out the lighting and correct the colors. They also came up with clever ways to make sure the checker looks right and works even if it's not perfectly placed in the image.
Why it matters?
This matters because it could make photos from all kinds of cameras look more accurate and natural without needing special training for each camera type. It could improve things like phone cameras, security cameras, and even help robots see colors more like humans do. This technology could make digital images more reliable and true-to-life across many different devices and situations.
Abstract
Color constancy methods often struggle to generalize across different camera sensors due to varying spectral sensitivities. We present GCC, which leverages diffusion models to inpaint color checkers into images for illumination estimation. Our key innovations include (1) a single-step deterministic inference approach that inpaints color checkers reflecting scene illumination, (2) a Laplacian decomposition technique that preserves checker structure while allowing illumination-dependent color adaptation, and (3) a mask-based data augmentation strategy for handling imprecise color checker annotations. GCC demonstrates superior robustness in cross-camera scenarios, achieving state-of-the-art worst-25% error rates of 5.15{\deg} and 4.32{\deg} in bi-directional evaluations. These results highlight our method's stability and generalization capability across different camera characteristics without requiring sensor-specific training, making it a versatile solution for real-world applications.