NVIDIA Posted Initial texts stylegan3 , machine learning systems based on a generative-sensitive neural network ( Gan ), aimed at synthesizing realistic images of people people. The code is written in Python using a Pytorch framework and distributed under the NVIDIA Source Code License license that imposes a limit on commercial purposes.
Downloads also Available Ready-made trained models trained in Flickr-Faces-HQ (FFHQ), including 70 thousand high-quality (1024×1024) PNG images of people people. In addition, there are models built on the basis of collections afhqv2 (pictures of animals) and Metfaces (images of people of people with portraits of classical painting). When developing emphasis is on the person, but the system can be trained to generate any objects, such as landscapes and cars. Additionally, the tools for self-study of the neural network on their own collections of images are provided. To work, one or more NVIDIA video cards is required (GPU Tesla V100 or A100 is recommended), at least 12 GB of RAM, PYTORCH 1.9 and CUDA 11.1+ toolkit. To determine the artificial nature of the received persons develops Special detector .
The system allows you to synthesize the image of a new person based on the interpolation of the characteristics of several persons, combining the features of them, as well as adapting the final image under the necessary age, the floor, the length of the hair, the character of the smile, the shape of the nose, the color of the skin, glasses, glasses. The generator examines the image as a collection of styles, automatically separates the characteristic parts (freckles, hair, glasses) from common high-level attributes (posture, gender, age-related changes) and allows them to combine them in arbitrary with the definition of dominant properties through the weight coefficients. As a result, images are generated, externally indistinguishable from these photos.
First version of the technology Stylegan