Dominik Kulon

Senior Machine Learning Engineer

Snap Inc. · London

Owner and lead developer of two core AR products: generative Body Avatars and 3D hand tracking.

• Video generation and model distillation (2024 – present). Developed Body Avatars, real-time on-device transformations of users into digital characters, by training and distilling pose-controlled human video diffusion models. Launched publicly in Lens Studio and EasyLens (Snap's AR creation platforms); individual effects reached over 300M views.
• 3D hand tracking (2020 – 2025). Built the hand tracking system that became a core component of AR effects, used by hundreds of millions of people and generating over 2B monthly views.
• End-to-end ownership. Own the full stack: architecture design, training strategies, evaluation, large-scale data collection and synthetic data generation, on-device optimisation, and production deployment.
• Impact-driven research. Evaluate models on engagement and product metrics, in addition to technical metrics, and use those signals to prioritise the research roadmap.

Computer Vision Scientist

Ariel AI · London Acquired by Snap Inc.

Early employee; joined during my PhD to bring research into production.

• Developed a real-time, on-device system for 3D hand mesh reconstruction from images.
• Built a 3D statistical model of human hands trained on 500k hand scans from 1,300+ individuals.
• Responsible for large-scale data collection and processing using a 3D motion capture setup.

Body Avatars

Snap Inc.

Links:Examples·Lens Studio guide·EasyLens Web

Developed a new category of augmented reality products that transform users into digital characters with pixel-level accuracy. The effects operate in real time on mobile devices; anyone can create their own character from a single reference image.

• Character consistency. Trained pose-controlled video models (Wan, Stable Diffusion) and image-editing LoRAs (Qwen-Image-Edit) so a character in a reference image stays consistent across poses.
• Distillation. Distilled video models into real-time on-device networks.
• Paired data generation. Built a pipeline that uses the pose-controlled model to produce image-aligned character annotations.
• Platform integration. Integrated the technology into Lens Studio and EasyLens (AR creation tools), enabling creators to train their own AR transformations.
• Agentic pipeline. Built an agentic concept design system that drives data generation, training, and publishing. It has produced thousands of effects, published to users and available to creators for remixing.
• Impact. The top five characters passed 1B views within a month of launch. Nearly 2k creators produced tens of thousands of AR effects.

Controllable Human Video Generation

Snap Inc.

Worked on pose-controlled human video diffusion, from improving gesture fidelity to 360° generation, 3D reconstruction, and digital character animation.

• Human video diffusion. Trained pose-controlled human video diffusion models, initially to improve hand and gesture fidelity. Identified an opportunity for digital character transformations that became the basis for Body Avatars.
• 360° generation and 3D Gaussian Splatting. Rendered a 360° human dataset, trained a video model to generate multi-view sequences from a single image, and reconstructed 3D humans with Gaussian Splatting from the generated views.

3D Hand Tracking

Snap Inc.

Links:Lens Studio guide

Developed a hand tracking system used by hundreds of millions of Snapchatters. Owned end-to-end product development, including model design and training, on-device optimisation, data collection, and production deployment.

• Modelling and training. Designed, built, and trained a 3D hand mesh reconstruction system.
• On-device optimisation. Developed low-end, mid-end, and high-end variants for different devices and use cases.
• Data collection. Ran data collection efforts to address failure cases. Created the semi-synthetic dataset for dense pixel-accurate supervision and differentiable rendering losses. Experimented with training a ControlNet-based data generator.
• Auxiliary models. Developed hand segmentation, surface normal, and depth estimation models.
• Impact. The system became a core component of augmented reality effects and enabled novel brand experiences for major companies such as Disney, Cartier, Dior, Lego, Netflix, Samsung, Coca-Cola, McDonald's, KFC, and many others.

Hand Shape Modelling

Ariel AI · Snap Inc.

Built a 3D statistical model of human hands and a registration system that brings hand scans into correspondence.

• Data collection. Ran large-scale 4D data collection, obtaining 500k scans from over 1,300 participants aged 3–81, with high-resolution meshes and textures.
• Registration. Built a pipeline for iterative optimisation that fits a 3D hand model to noisy, incomplete scans.
• Statistical deformable model. Trained a 3D hand shape model with independent bone-length and shape variations.
• Synthetic dataset. Trained a texture model and built a semi-synthetic dataset with normal, depth, segmentation, and UV annotations.