This work presents an effective deep learning-based translation approach between photo-realistic images and stylized cartoon representations, using a further improved version of the CartoonGAN network. It is based on a generator–discriminator model; the generator consists of an encoder–residual–decoder structure to map input photos to cartoon-style outputs, while the discriminator is responsible for ensuring that real cartoon images are differentiated from both generated and smoothed references. Image input validation, resizing, and normalization are performed in consideration of the proposed preprocessing steps. Adversarial losses, along with content losses computed from intermediate VGG-16 features, are used for training, with maintenance of structural fidelity and cartoon realism. For model optimization, Adam is used with tuned hyperparameters. Results were obtained by conducting qualitative and quantitative analyses. The generated images retained sharp edges, minimal texture, and a visually appealing appearance in their cartoon form. Among the metrics evaluated, the peak signal-to-noise ratio (PSNR) stood at 19.7681 dB, which essentially represents strong reconstruction quality with visual fidelity of the cartoonized output of the model.