Supplmentary Material: L-CAD: Language-based Colorization with Any-level Descriptions using Diffusion Priors

To demonstrate the effectiveness of our proposed luminance-guided image compression, semantic-aligned latent representation, and instance-aware sampling strategy (details in Sec. We demonstrate our generalization capability by showing more colorization results on legacy black-and-white photos in Figure 1, where results are presented sequentially from left to right using descriptions at the complete, partial, and scarce levels. Learning to color from language.