Taras Khakhulin

I am research engineer at at Synthesia , where we’re working on the future of media 🚀. My current focus is on developing generative models based on diffusion processes for video and motion. Before I was a Ph.D. student at Skoltech advised by Victor Lempitsky, and worked as a research engineer at Samsung AI Center for several years, where we pushed the boundaries of view synthesis. I try to synthesize things that people currently cannot. During my studies, I have contributed to 3D representations, image synthesis, and human avatars.

Prior to that, I investigated the application of reinforcement learning for discrete optimization under the advising of Ivan Oseledets. Earlier, I was a part of the DeepPavlov project, where we studied word representations for text with missplellings using context-aware language modeling.

Email  /  CV  /  Google Scholar  /  Twitter  /  LinkedIn  /  Github

Introduced high-resolution dataset of human performances to speed up development in avatars. Also, we demonstrated ability to reconstruct large scenes with temporal factorization for 4D NeRFs.

Create the most efficient renderable representtion for novel view synthesis from an arbitrary number of images (more than one). We learn this system end-to-end on the dataset of monocular videos. Extremely fast and compact way to present scene.

Create an animatable avatar just from a single image with coarse hair mesh and neural rendering. We learn head geometry and rendering together with supreme quality in a cross-person reenactment.

One-shot high-resolution avatars with latent poses for cross-reenactment.

The scene can be represented as a set of semi-transparent mesh layers from just stereo pair without loss of the quality. This representation allows effortless estimation and fast rendering. Additionaly, we published a dataset with occluded region - SWORD - for novel view synthesis.

Our generator produce images without any spatial convolutions. Each pixel synthesized separetly conditioned on noise vector. We investigate properties of such generator and propose several applications (e.g. super-res, foveated rendering).

We propose a learning heuristic with RL for real-world combinatorial problem on graphs. Suprisingly, we can easily estimate universal policy which can be scaled across different graphs.

Our generator produce images without any spatial convolutions. Each pixel synthesized separetly conditioned on noise vector.

The heuristic approach based on tree decomposition to relax the contraction of tensor networks using probabilistic graphical models and applied for random quantum circuits.

We suggest a new language-independent architecture of robust word vectors (RoVe). It is designed to alleviate the issue of typos, which are common in almost any user-generated content, and hinder automatic text processing. Our model allows it to deal with unseen word forms in morphologically rich languages before contextualized language models.

Open-source library for end2end conversational NLP.

The webpage template was borrowed from Jon Barron.