Daniel Geng

I am a fifth year PhD student in CS at the University of Michigan advised by Professor Andrew Owens. I study computer vision, and have previously done research in deep reinforcement learning and representation learning.

I spent last summer as a student researcher with Charles Herrmann and Deqing Sun at Google DeepMind. In the past I was lucky enough to have worked as an undergrad at UC Berkeley under Sergey Levine and Coline Devin, as well as Alyosha Efros and Taesung Park, and as an intern at FAIR under Lorenzo Torresani and Huiyu Wang.

I am currently looking for full-time jobs!

Github  /  Google Scholar  /  Twitter  /  Email

I am currently interested in generative models, how we can control them, and novel ways to use them. I've also worked on levarging differentiable models of motion for image and video synthesis and understanding, and in the past I've worked on representation learning, multimodal learning, and reinforcement learning.

We train a video generation to be conditioned on motion, and then prompt it with motion prompts, elliciting a wide range of behavior.

We use diffusion models to make spectrograms that look like natural images, but also sound like natural sounds.

Sister project of "Visual Anagrams." Another zero-shot method for making more types of optical illusions with diffusion models, with connections to spatial and composition control of diffusion models, and inverse problems.

A simple, zero-shot method to synthesize optical illusions from diffusion models. We introduce Visual Anagrams—images that change appearance under a permutation of pixels.

We achieve diffusion guidance through off-the-shelf optical flow networks. This enables zero-shot motion based image editing.

By differentiating through off-the-shelf optical flow networks we can train motion magnification models in a fully self-supervised manner.

Pixelwise losses compare pixels by absolute location. Instead, comparing pixels to their semantic correspondences surprisingly yields better results.

Life seeks order. If we reward an agent for stability do we also get interesting emergent behavior?

Learning a composable representation of tasks aids in long-horizon generalization of a goal-conditioned policy.

Given any classification architecture, we can augment it with a confidence network that outputs calibrated class probabilities.

Website template from Jon Barron.
Last updated Feb 2024.