Onur Tasar

Onur Tasar

Research Scientist · Paris, France

I develop AI systems built on multimodal foundation models to solve challenging real-world problems. My work spans problem formulation, model design, large-scale training, rigorous evaluation for production deployment.

Over the past few years, I have developed and shipped generative AI systems for image editing and generation that are used by millions of users and generate seven-figure annual revenue.

I currently lead the Image Research team at Jasper, developing state-of-the-art generative and multimodal models. Previously, I was a Research Scientist at Stability AI. I hold a PhD in Computer Vision and Machine Learning from Inria.

Interests Generative image editing · Diffusion models · Multimodal learning · Distributed training · Synthetic data

Onur Tasar

Selected research & projects

Applied research work that I directly contributed to, spanning research, code, method development, and training at scale.

Image expansion (Uncrop) result

Image expansion (Uncrop)

An outpainting system that extends images beyond their borders with coherent results, finetuned from Flux with a ControlNet.

Product

Controllable shadow and reflection synthesis

Controllable shadow & reflection synthesis

A diffusion-based model with dedicated conditioning mechanisms for realistic, controllable shadows and reflections in product photography, released with an accompanying benchmark. Users can control shadow direction, softness, and intensity.

Project page Benchmark Product

Native RGBA object removal

Native RGBA object removal

Diffusion-based editing with a custom RGBA variational autoencoder (VAE) that jointly models transparency alongside standard color channels. This native RGBA support unlocks advanced, controllable workflows for design and content creation.

Write-up Product

Blender-based synthetic data generation pipeline

Synthetic data generation pipeline

Built a rendering pipeline from scratch on top of Blender to generate perfectly annotated, large-scale datasets for training multimodal generative models. It powers the training of several of our models, including object removal, relighting, and shadow and reflection prediction.

Dataset

High-fidelity face swap

High-fidelity face swap

Trained a custom SDXL model allowing the users to replace the face of a source image with the face of a target image.

Product

Real-time image relighting

Real-time image relighting

Generative relighting of images and objects in real time, with interactive control over the number of lights and their properties, including color, intensity, and position. This tool went viral on various social media platforms with millions of views and likes.

Product Technicalities A user video with 700k+ likes

Selected publications

During my PhD I published 4 journal papers, 5 conference papers, and a book chapter (700+ citations), reviewed for top venues such as ICCV, and co-supervised a master's intern whose internship concluded with a publication.

Experience

Contact

Feel free to reach out to me via the links below.