Andrey Voynov

I'm a researcher at Google in Tel Aviv, focusing on creative imagenery.

I work in Creative Camera focusing on new capabilities of generative models. I also host visiting student researchers and collaborate with academic groups. I have math background and defeated my PhD from Moscow State University in 2014. Before joining Google in 2022 I worked as a Research Scientist in Yandex Research and also participated in its autonomouse car developement.

Email  /  Scholar  /  𝕏  /  Github

Research

I'm interested in computer vision and deep learning, in particular, images generative models and unsupervised learning. My current research is mostly focused on new capabilities of visual generative models for creativity. My math research was in the intersection of convex geometry and functional analysis.

A diffusion model personalization is performed with the prior knowledge of the target prompt to be used for.

Euler forward method is used to enable more accurate diffusion inversion.

Diffusion model with extra camera curvature conditioning implemented with either Riemannian metric tensor, or per-pixel coordinates conditioning.

Cross-batch shared self-attention makes a diffusion model generate images with aligned styles.

Diffusion model personalization forms a binary tree of a visual concept decomposition.

Small MLP performs gradient guidance over intermediate diffusion features for sketch-to-image generation.

Different prompts are injected to different cross-attention layers that majorly improves textual inversion and allows appearance mixing.

Recall is what important for GAN initialization, and Imagenet-pretrained StyleGAN is a good choice.

Diffusion model intermediate features are used for few-shot segmentation.

Background-segmentation latent direction of BigBiGAN produces synthetic data for unsupervised foreground segmentation learning.

Finding StyleGAN weights shifts that induces interpretable images editing.

Self-supervised pretrained backbones are shown to be better features extractors for GANs evaluation.

An unsupervised method to find interpretable directions in a GAN latent space.

A GAN with a generator composed of a sequence of randomly-chosen layers.

Math Papers

My math research was primarly focused on functional analysis, random matrices semigroups, and convex geometry. I had a pleasure to have Vladimir Protasov as my PhD advisor. In all the papers below the authors order is alphabetical.

This page template is based on Jon Barron's public academic website source code.