Description
We are developing Russia's first World Model based on the Kandinsky video generation model. Input: description, frame, script, or a set of actions/instructions; output: an interactive scene. The focus is on developing new architectures and techniques, training large models (tens/hundreds of billions of parameters), and optimizing inference.
Responsibilities
- research and development in the field of video generation and world models (researching existing architectures and developing new ones)
- work on various aspects of project activities: reading papers and formulating hypotheses, planning and setting up experiments, including data collection and preparation (if required), processing and analyzing results, defending proof of concept, presenting results at reporting events, writing papers and technical documentation
- work on topics such as improving the physical realism of generation, measuring and enhancing spatiotemporal consistency, integrating control context into generation, various types of optimization aimed at increasing autoregression stability and accelerating generation, and other world model modules
- collaborative work both within the team building the World Model and with other teams responsible for data collection and preparation, pre-training, post-training, optimization, open-sourcing, production, etc.
- interaction with partner teams for model adaptation and integration (autonomous vehicles, robotics, video game industry).
Requirements
- expert level Python, PyTorch
- deep understanding of ML/DL/CV and visual GenAI
- experience in classical image and video processing tasks
- experience in processing large video datasets
- experience with diffusion models
- understanding of training/distributed training methods
- understanding of modern LLM and Diffusion model architectures
- understanding of video quality assessment metrics.
Nice to have:
- understanding of 3D and its relation to image/video (point clouds, depth maps, voxels, mesh, NeRF, Gaussian splats, etc.)
- understanding of RL principles
- understanding of digital signal processing, compression, and enhancers
- understanding of the concept of autonomous vehicles or robotics
- ability to explain complex things in simple terms.
Conditions
- comfortable modern office near Kutuzovskaya metro station
- hybrid work format
- annual salary review, quarterly and annual bonuses
- corporate gym and relaxation areas
- over 400 SberUniversity educational programs for professional and career development
- onboarding program and manager's support at the start
- extended voluntary health insurance, preferential insurance for family, and corporate pension program
- mortgage benefits up to 7% for every employee
- free SberPrime+ subscription, discounts on products from partner companies
- referral bonus for recommending friends to join the Sber team.
