A Chinese Ink-Wash Style Transfer Method for 3D Scenes

Rendering 3D scenes in traditional Chinese ink-wash style has been a topic worth researching. While current 3D ink-wash style rendering methods can produce satisfying static images that reproduce the artistic style of ink-wash paintings, they often lack versatility and fail to deliver dynamic visual effects in certain circumstances. We propose an ink-wash style transfer method for 3D scenes that leverages neural style transfer techniques primarily targeted for images. To begin, we use a landscape ink-wash painting as the style reference to train a style transfer model and transfer ink-wash style presented by the style reference onto given input image, while retaining its original content. Subsequently, we apply this model into a real-time environment and stylize 3D scene objects in the trained ink-wash style. During the training of our model, Sliced Wasserstein loss is leveraged for enhanced style representation and depth loss is introduced to further emphasize depth information of objects. Qualitative and quantitative experiments showed that our approach achieved more appealing visual effects and better performance compared to state-of-the-art methods.