Yotam Nitzan

I am a Research Scientist at Adobe in San Francisco.

Previously, I got my Ph.D. student in Computer Science at Tel-Aviv University, where I was advised by Prof. Daniel Cohen-Or. I've interened at Google Research and Adobe Research (x2). Before that, I received M.Sc. in Computer Science from Tel-Aviv Univeristy and B.Sc. in Applied Mathematics from Bar-Ilan University.

I'm focused on visual generative models, trying to make them more interactive, intuitive, and controllabe.

Email  /  Github  /  Google Scholar  /  Twitter

LazyDiffusion is an efficient image editing architecture that updates only user-specified regions. Using a context encoder and a diffusion-based transformer decoder, it balances global context awareness with localized generation, achieving state-of-the-art quality and up to 10× speed improvements.

We present a method to expand the generated domain of a pretrained generator while respecting its existing structure and knowledge. To this end, we identify dormant regions of the model's latent space and repurpose only them to model new concepts.

We survey the family of StyleGAN models and how they have been employed for downstream applications since their inception.

We present the first personalized face generator, trained for a specific individual from ~100 of their images. The model now holds a personalized prior, faithfully representing their unique appearance. We then leverage the personalized prior to solve a number of ill-posed image enhancement and editing tasks.

Two models as considered aligned if they share the same architecture, and one of them (the child) is obtained from the other (the parent) via fine-tuning to another domain, a common practice in transfer learning. In this paper, we perform an in-depth study of the properties and applications of aligned generative models.

We propose a novel method for solving regression tasks using few-shot or weak supervision. At the core of our method is the observation that the distance of a latent code from a semantic hyperplane is roughly linearly correlated with the magnitude of the said semantic property in the image corresponding to the latent code.

We identify the existence of distortion-editability and distortion-perception tradeoffs within the StyleGAN latent space on inverted images. Accordingly, we suggest two principles for designing encoders that are suitable for facilitating editing on real images by balancing these tradeoffs.

A generic image-to-image framework, based on an encoder that directly maps into the latent space of a pretrained generator, StyleGAN2.

We propose to disentangle identity from other facial attributes by mapping directly into the latent space of a pretrained generator, StyleGAN.

Website template courtesy of Jon Barron.