StyleGAN
This is a graphical interface for editing portraits. The basis is the ArcFace model. On the screen there are fields for entering text, and with the help of keywords you can determine editing parameters: change color, volume and length of hair, remove or add makeup, and even easily adjust the age of the character.
EigenGAN
The system itself identifies hidden spaces and uses them for manipulations: changing the character's gender, rotating the body, changing pose or hairstyle. It only struggles with glasses because they are rare in the data, and sometimes confuses gender or pose. Other features are handled well.
ReStyle
To edit an image, the model inverts the image's latent code. Instead of a single pass to predict the code, the system calculates at each step the residual relative to the current state of the inverted latent code, so quality improves significantly.
Geometry-Free View Synthesis
The system builds a three-dimensional image from a single photograph. It's enough to upload an image of a room or part of an apartment, and it will complete several variants itself. A quantized space representation is used, without the need for ready-made 3D models or geometry descriptions – the system learns spatial parameters on its own.
LatentCLR
Works with the latent space of GAN models and identifies meaningful vectors. Uses contrastive learning without human supervision. Non-linear vectors are identified in trained versions like BigGAN and StyleGAN2.
Articulated Animation
Capable of creating full-body deepfake: separates the body from the background, identifies movement style and generates new movements. Not dependent on a specific person – learns and applies to anyone you input.
VideoGPT
New architecture for video generation. Uses VQ-VAE automatic encoder to create a latent representation of video without labeling, with three-dimensional convolutions and self-attention, and then GPT for autoregression and time encoding.
MiVOS
Tracks objects in video and creates binary masks. Masks are transferred between frames thanks to a convolutional network, and the user can compare and correct at any point through a convenient graphical interface.
DINO
Innovative approach without manual labeling: combination of transformers and self-learning. Models learn on unlabeled data, apply selective focus and generate hypotheses.
CPA
Predicts the effect of combinations of connected features.
