Beyond the Lines: Grayscale-Driven Dual-Guided GANs for Enhanced Sketch Image Colorization

Seunghoo Lee, Junehyoung Kwon, Jooweon Choi, Mihyeon Kim, Eunju Lee, YoungBin Kim

Abstract: In this paper, we present a novel Dual-Guided Colorization (DGC) framework designed to enhance the per-formance of sketch image colorization models. By leveraging grayscale images, which provide richer information such as texture and luminance, the DGC framework effectively reduces color bleeding, a common issue in sketch-based colorization tasks. We also introduce the Color Bleed Index (CBI), a new metric for quantifying and visualizing color bleeding. Experimental results on the AnimeDiffusion dataset demonstrate significant improvements in key metrics, including Fréchet Inception Distance (FID), Structural Similarity Index Measure (SSIM), and CBI, confirming the effectiveness of our approach. Notably, DGC improved FID by up to 10.88%, SSIM by 1.49%, and CBI by 8.15% compared to baseline methods.