<![CDATA[Photo-to-cartoon translation is a non-photorealistic rendering task that generates illustrative visuals while preserving fundamental object structures. This study proposes a CartoonGAN-based approach employing a joint learning scheme that integrates a lightweight denoising module into the generator. Trained end-to-end alongside the stylization process, this module suppresses noise and irrelevant textures without losing critical semantic information from input photographs. Using unpaired photo and cartoon images from the Hugging Face platform, the model is trained with a combination of adversarial and L1-based content losses to balance style generation and structural preservation. Experimental results indicate a stable and convergent training process, achieving an average content loss of 0.0286 and a generator adversarial loss of 0.3982 at epoch 50. Qualitatively, the generated images exhibit sharper contours, uniform color regions, and reduced fine textures compared to the original photographs. These findings demonstrate that integrating a denoising module via joint learning significantly improves visual consistency and training stability, providing an effective deep learning-based solution for photo-to-cartoon translation.]]>
Copyrights © 2026
