Alibaba’s AI Research Team Introduces “DCT-Net,” a new image translation architecture for few-shot portrait styling.
Portrait stylization tries to change how a person looks into more creative styles (like the 3D cartoon, anime, and hand-drawn) while keeping the person’s identity. This research topic has a lot of uses in the digital arts, such as making art, making animations, and making virtual avatars. Unfortunately, these works of art can be by people with specific creative skills and a lot of hard work.
In scientific literature, there are two main ways to change a person’s picture style automatically. First, image-to-image translation techniques automatically learn a function that maps an image from the source domain to the target domain. But these methods need a lot of data and cause texture artefacts when the scene in the picture is complicated. On the other hand, most recent works use a pre-trained StyleGAN to stylize portraits, but we can’t use it to make photographs look like other portraits. Also, these methods only work with head images and can’t deal with full-body images.
The DCT-Net framework’s objective is to discover a function that converts images from the source domain into those of the destination domain. While maintaining the source image’s content features, the image produced by the framework should display the target domain’s texture style. In addition, the GEM applies geometric adjustments to the source and target domain samples to facilitate the production of full-body pictures. The TTN then gains knowledge on how to produce a cross-domain translation. The CCN and the TTN are trained separately during the training procedure. Furthermore, following training, DCT-Net exclusively uses the TTN for its concluding inferences. In the same way, the goal of the CCN is to set the distribution of the few samples from the target domain.
Conclusion
The researchers showed DCT-Net, a new framework for making stylized portraits. DCT-Net improves the head stylization task regarding ability, generality, and scalability. It also does an excellent job of translating the whole body image elegantly. The main idea is to start by calibrating the biased target domain and then learn a fine-grained translation.
Their experiments showed that their method was better and worked better. The researchers also thought that their solution of domain-calibrated translation could lead to more research on image-to-image translation tasks with biased target distribution.
Source: indiaai.gov.in
