Digital film post-processing, a crucial step in film production, demands higher image quality and processing efficiency. However, traditional techniques struggle with low computational efficiency and insufficient adaptability when handling high-resolution and complex dynamic scenes. In recent years, the rapid development of deep learning technology has opened up new possibilities for addressing these issues. This research constructs an improved processing model based on a convolutional neural network (CNN), combined with the RMSprop optimization algorithm and pruning techniques, to enhance the model’s feature extraction capabilities, convergence speed, and real-time application performance. Experimental results demonstrate significant advantages of the improved algorithm in image processing tasks. In the denoising task, the Peak Signal-to-Noise Ratio (PSNR) increased from 26.4 dB with the traditional method to 34.8 dB, with an average processing time reduction of approximately 21.5%. In the real-time special effects synthesis task, the processing time of the improved algorithm was shortened to 31.8 ms, and the GPU memory usage was reduced by about 15%. User experience evaluations showed scores of 8.7, 8.9, and 8.6 for realism, detail preservation, and smoothness, respectively, representing over a 30% improvement compared to traditional methods. In summary, the improved CNN algorithm, through optimized learning strategies and resource utilization, not only enhances image processing quality but also meets real-time application requirements, providing an efficient technical solution and theoretical support for digital film post-processing.