Language-Guided Colorization: Reconvolution and Cross-Modal Align

Previous language-guided colorization methods have obvious flaws, like color spillover and inconsistent themes. The root lies in poor info transfer and feature processing in models. Redundant data in transfer lowers efficiency and accuracy of feature representation. Existing feature-fusion modules also can’t integrate diverse features well, harming colorization. We propose a new approach with two key modules. The Reconstruction Convolution Module (RCM) cuts costs by removing redundant info and strengthens feature representation. The Cross-Modal Color Aligner Module (CMCAM) uses multi-scale features to align color and grayscale features precisely, improving semantic understanding. Experiments show our method outperforms state-of-the-art, achieving better results and robustness in high-quality color image generation.