Neta Shaul

I am a PhD student at the Weizmann Institute of Science under the supervision of Yaron Lipman.

Email  /  Scholar  /  Twitter

Research

My research focuses on developing and advancing scalable modeling frameworks for generative models. I'm interested in a variaty of data types from both discrete and continuous domains (text, images, videos, proteins, etc).

Transition matching (TM) replaces the infinitesimal-timestep kernels from Flow Matching/Diffusion with a generative model, advancing both flow/diffusion and autoregressive models. TM variants achieve state-of-the-art text-to-image generation.

Through the lens of kinetic optimality, we expand the design space of Discrete Flow Matching, allowing the use of any probability path and simultaneously justifying existing mixture paths.

Flow Matching for discrete data with applications to language modeling and image generation.

Optimizing model-specific non-stationary ODE solver for fast sampling of pre-trained Diffusion and Flow-Matching models.

A novel method to build costume-made ODE solvers for sampling pre-trained diffusion/flow models. Significantly improves generation quality for low number of function evaluations.

We study Kinetic Optimal paths within the class of probability paths used for training generative models, and show that in high dimension Cond-OT path becomes the kinetic optimal.

Design and source code from Jon Barron's website.