'Hyper Style' that allows you to change the input image to your favorite style in near real time
'StyleGAN' developed by NVIDIA is a technology that generates high-quality images based on the learned data, and is being improved steadily. A research team at Tel Aviv University has refined and trained the weighting of 'reconstructibility' and 'editability' in the StyleGAN generator to efficiently re-learn a single input image in less than two seconds. I created the 'Hyper Style' that composes.
HyperStyle
[2111.15666] HyperStyle: StyleGAN Inversion with HyperNetworks for Real Image Editing
GitHub --yuval-alaluf / hyperstyle: Official Implementation for 'HyperStyle: StyleGAN Inversion with HyperNetworks for Real Image Editing' https://arxiv.org/abs/2111.15666
https://github.com/yuval-alaluf/hyperstyle
For example, in the photo of actor Brad Pitt shown below, the one on the left is the original one. The image on the right is an image generated by converting the input image to a latent space vector and inputting it to StyleGAN.
It is known that the output image changes when the variable of the latent space vector is tampered with. For example, you can change the hairstyle or age based on the input image as shown below. ..
HyperStyle speeds up image reconstruction. Click the image below to see how the input image (left side) is immediately reflected in the generated image (right side).
There are also movies that have been tested with not only pictures of people's faces but also pictures of cars. In this movie, you can clearly see that there are differences from the input image in small parts such as the color of the car body and the license plate.
The following is a movie in which the smile, beard, and hairstyle are changed while retaining the facial features of the input image. The sample is Leonardo DiCaprio, which looks different without any discomfort.
Furthermore, it can be changed to an anime style or a painting style.
According to the research team, it can be generalized to other than the images generated by StyleGAN.
Related Posts:
