CoGS: Controllable Generation and Search from Sketch and Style abs: arxiv.org/abs/2203.09554

Some pictures of tonight's lunar eclipse. Taken with a cell phone mounted to an 80 mm refractor.

Excited to share "CoGS: Controllable Generation and Search from Sketch and Style" at #ECCV2022! (w/ Gemma Canet Tarrés, Tu Bui, James Hays, John Collomosse, Zhe Lin) Project page: cusuh.github.io/CoGS/

James Hays is our guest lecturer at I2DL this semester! TU München Starting with the awesome Argoverse2 dataset! @argoai

Working on a walkthrough of transformer code with side-by-side comparison to the computation graph. Felt the default image everyone uses is hard to interpret. Also, I like to show the matrices in my computation graphs. § refers to sections in a colab notebook (coming soon)

Computing is the fastest growing College at Georgia Tech. So of course my home unit, the School of Interactive Computing, is hiring faculty. Stay tuned for a proper ad. For now, I polled our faculty and came up with this unofficial list of priority growth research areas.

I continued exploring #stablediffusion's latent space over the weekend and oh my; there's still a LOT of treasure to be discovered inside this magnificent neural universe! Here's a quick thread with some of my personal favorites and how I found them..

Argoverse 2: Next Generation Datasets for Self-Driving Perception and Forecasting abs: arxiv.org/abs/2301.00493

Augmentations are important in ML, but did you know a soft augmentation is the way to go? Inspired by human visual classification studies, the more aggressive your image crop augmentations, the more your learning target softens. From my time at Argo! arxiv.org/pdf/2211.04625…

Thank you to my collaborators (James Hays, Cynthia, Krishna, Zhifei, and Tobias)! Machine Learning at Georgia Tech Adobe Research

Our work ZeroFlow: Scalable Scene Flow via Distillation is accepted at #ICLR2024! We show that scaled up distillation from a noisy teacher on unlabeled data outperforms human supervision to capture SotA! 🌎Website: vedder.io/zeroflow 📄Paper: arxiv.org/abs/2305.10424

Scene Flow methods promise to describe motion on any moving object… …but they can't even describe pedestrians walking! To focus the community on this problem, we're announcing the Argoverse 2 2024 Scene Flow Challenge focused on this issue: argoverse.org/sceneflow 🧵👇

Scene Flow promises to be a catch all solution for open-world detection of moving objects, but in practice they fail on the second most common class in AV datasets pedestrians! Solving this is the goal of this year's Argoverse 2 Scene Flow challenge! argoverse.org/sceneflow 🧵👇

In our latest work, we introduce personalized residuals for efficient personalization of T2I diffusion models and localized attention-guided sampling that dynamically blends the learned concept into new contexts. #CVPR2024 cusuh.github.io/personalized-r…

I'll be presenting our work tomorrow (6/19) 5-630pm in Arch 4A-E (poster #329)! #CVPR2024

We just hosted a Student Conference Machine Learning at Georgia Tech, bringing together students/fac from across campus to build a stronger community. Grateful to all participants & proud of the incredible work showcased! It was a success cuz we ran out of food :) Hear our Atlanta AI Buzz Georgia Tech!🐝

Fascinating to be reminded that image manipulation dates back to the 1800s, especially in today’s generative AI era. Big thanks to Walter Scheirer for an insightful talk at Machine Learning at Georgia Tech ! Check out his new book on the history of fake images and their social impact: sup.org/books/title/?i…