Region Weighted Conditional GAN for Human Skin-Focused Grayscale Image Colorization

This work addresses a failure mode that generic colorizers share but do not treat: regions whose plausible colors are narrowly constrained. Human skin admits only a small palette, yet in our experiments it was the region most often colored unrealistically or left uncolored. We introduce a region-weighted reconstruction loss that adds a masked L1 term over the skin region alongside the standard full-image L1 and adversarial terms, with the total reconstruction weight held constant so that any gain is attributable to the region term rather than to a shifted loss balance. We also contribute a skin mask derivation from the CIHP parsing annotations and a 60,011-image training set that combines CIHP with skin-free Places 365 images, so the model sees both skin and non-skin scenes. Evaluated within the skin region across two generator architectures, the region term improves PSNR, MAE, and SSIM in both pairs, with the best model reaching PSNR 36.98, MAE 1.22, and SSIM 0.99, and receiving the highest ratings in a blind human review. A secondary finding is that FID and Inception Score mildly favor the unfocused models even as human reviewers prefer the focused ones, indicating that distribution-level metrics do not capture perceptual gains confined to a small region.