Happy to share that our paper has been accepted to #CVPR2024. arxiv.org/abs/2312.07330 Huge thanks to our collaborators!

Computer Vision conference's acceptance criteria these days: #CVPR2024 #eccc2024 #AI #ComputerVision

Join us this Wednesday, 10:30-12:30 at #ECCV2024, Poster Session 3, #173 for our ♾-Brush🖌️paper! Also, my advisor Dimitris has 8 ECCV papers this year. Feel free to discuss diverse topics from generative models, Gaussian splatting to human attention, segmentation!

🚀 Introducing Hummingbird 🐦: High-Fidelity Image Generation via Multimodal Context Alignment! 🎨✨ We’re excited to announce Hummingbird’s acceptance to ICLR 2025! 🎉 Developed during my Microsoft internship, it explores new approaches to high-fidelity image generation.

Cool idea using DMD for post-training - elegant to apply differentiable losses directly without RL. In our Hummingbird (ICLR’25), we take a similar route: direct backprop instead of RL, generating diverse images w.r.t. a reference while maintaining key properties via

Physics enters video generation. Excited to share our work NewtonRewards: the first framework that enforces Newton’s Laws in video diffusion models using verifiable rewards-no labels, no human feedback. Stronger physics, smoother motion, real dynamics. 🔗 cvlab-stonybrook.github.io/NewtonRewards

The paper shows how to teach video generators basic gravity rules so objects move more like in real life. NewtonRewards is a post training method that scores generated videos by how well they obey simple Newtonian motion rules. Standard video diffusion models mainly mimic

Video generation models still struggle with basic physics - objects float, drift, or accelerate incorrectly. NewtonRewards changes that. We derive a Newtonian kinematic constraint that enforces constant acceleration across 5 Newtonian Motion Primitives, and pair it with a