Bimodal Neural Style Transfer for Image Generation Based on Text Prompts

Neural networks have become one of the essential areas in Artificial Intelligence due to their extraordinary capacity to address problems in different domains. This ability led to the proposal of novel architectures and models to tackle challenging tasks such as neural style transfer. We propose a novel methodology for bimodal style transfer using text as input. We initially retrieve one image and a short descriptive text, which are mapped into a multimodal common latent space. Then, a new image is retrieved using an image retrieval engine. Finally, we use a generative model, which allows us to create artistic images by combining content and style. The proposed system can retrieve semantically similar images concerning a descriptive text (prompt), achieving great precision rates in image retrieval applied to the SemArt dataset. The transfer style neural model also preserves the image's high quality, combining style and content.