In minimally invasive surgery, endoscopic imaging is often compromised by insufficient illumination, which can degrade image quality and increase both the difficulty and risk of surgical procedures. In recent years, denoising diffusion models have shown considerable promise for enhancing low-light images, demonstrating strong capabilities in detail restoration and noise suppression. However, their high computational demands and limited ability to model low-dimensional features restrict their suitability for real-time clinical use.To overcome these limitations, we propose a lightweight hybrid diffusion framework, termed CTDiff, which combines a generator and a discriminator. Within the generator, local and global low-dimensional features are first extracted and fused through the convolutional neural network and the Swin Transformer, enabling initial enhancement of low-light images. This is followed by the iterative inverse diffusion process to refine residual noise and blurred regions, thereby further improving image quality. A dynamic resolution strategy is also introduced during generation to reduce computational overhead. In the discriminator, the global–local discriminative mechanism is employed to regulate brightness distribution and maintain structural consistency, enhancing both the realism and stability of the enhanced outputs. Experiments conducted on the simulated low-light endoscopy dataset demonstrate that the proposed method achieves notable improvements in brightness restoration and detail preservation. Moreover, the model enhances the accuracy of downstream surgical instrument segmentation, underscoring its potential for clinical application.