Influence of Color Space on GAN-Based Image Colorization with a Grayscale to Color Network

This work isolates a variable that the colorization literature has consistently left uncontrolled. Prior GAN-based colorizers each pick a color space, but the choice is rarely justified, and the few papers that do compare spaces vary the architecture at the same time, so the effect cannot be attributed. We hold the cGAN generator, discriminator, loss function, optimizer, and training schedule fixed and vary only the color space across RGB, Lab, HSV, YUV, and YCbCr. The result is a clear and reproducible trade-off: YCbCr yields colors closest to the ground truth (PSNR 21.29, MAE 16.42) while Lab yields the structurally smoothest output (SSIM 0.87), and a blind review with 78 participants confirms both as the preferred spaces. We then apply this finding to select a generic colorizer, showing that a ResNet-50 encoder generator trained in Lab under a generator-ahead schedule reaches PSNR 22.38, SSIM 0.90, and FID 45.88 on DIV2K, competitive with published methods. The contribution is therefore a controlled empirical basis for a design decision that is currently made by convention.