Colour Style Generation for Digital Images: A Controllable Diffusion Model with Semantic Segmentation Constraints
Y. Q. Chen, H. L. Du · Advanced Electromagnetics · 2026
Artificial intelligence generation techniques are increasingly reshaping digital image colour editing by enabling flexible and controllable style reconstruction. A controllable diffusion model with semantic segmentation constraints is proposed to address semantic mismatch, boundary colour bleeding, and insufficient regional control in global colour grading. Built on a latent diffusion framework, the model integrates semantic segmentation, multi-scale regional feature encoding, Lab-space colour statistics, style condition embedding, cross-attention injection, semantic-mask-guided denoising, and multi-objective consistency optimisation. Region-specific colour constraints are dynamically introduced during reverse diffusion to preserve object structure, coordinate local tones, and support continuous adjustment of style intensity, hue, and saturation. Experiments show that the proposed model achieves an FID of 23.7, 19.4% lower than ControlNet, while SSIM reaches 0.891. Compared with ControlNet, LPIPS decreases from 0.171 to 0.142 and colour error ∆E drops from 8.2 to 6.8. Semantic mIoU increases to 82.6%, boundary F1 reaches 87.4%, and style similarity improves to 0.913, confirming superior generation quality, semantic preservation, and controllability. Ablation results verify complementary contributions from segmentation, embedding, consistency, and boundary smoothing.