Vasu Singla

Thank you for the retweet! Our dataset, PixelProse contains descriptive and dense captions for over 16M images through the Google Gemini Vision model! We carefully curate images from 3 different sources, filter for CSAM, and provide additional filters and metadata.

⛷️Here’s my entry for the fast generative model olympics🥇 The Sphere Encoder is an autocoder so powerful that it produces high quality images quickly and without diffusion. At training time, we learn an encoder that maps natural images uniformly onto the surface of a sphere. At inference time, we sample a random vector from the sphere, and a decoder makes it into an image.

@jbhuang0604 @Songwei_Ge Congratulations, Songwei!

Proud advisor moment 😊 Congrats @Songwei_Ge for winning the Larry S. Davis Doctoral Dissertation Award @umdcs! Songwei is now cooking as a research scientist at @reve. Looking forward to amazing work!

Guardrails with custom polices are hard for models trained on safety and harm-related datasets. But what if you trained a guardian model on arbitrary rules? Introducing DynaGuard, a guardian model for custom policies: arxiv.org/abs/2509.02563

@ahans30 Congrats Abhimanyu!

zoom bombing is lame guys, especially in 2025, especially in someone's PhD proposal talk totally unrelated but guess who's a PhD candidate now 👀

@imVinusankars @ml_umd @FeiziSoheil @dmanocha @abhi2610 @HernisaKacorri @umdcs Congratulations Vinu!

📢Exciting! I have successfully completed my #PhD from @ml_umd on AI Safety🥳 Grateful to my family, teachers, and friends😇 Special thanks to the most supportive advisor @FeiziSoheil and committee members! Thank you to all those who attended my defense and to my collaborators❣️

@BasuSamyadeep @Adobe Congratulations Samyadeep!

Excited to start as a Research Scientist at @Adobe after a great time at UMD! I am going to be working on topics in language model reasoning and multimodality. Reach out if you are interested to collaborate!

Introducing MORSE-500 🌐 morse-500.github.io 500 scripted videos that stress-test six reasoning skills — beyond math, beyond static pics, built to get harder. Key Features: 🚀 Fresh & Portable 🎯 Diverse Categories 👁️ Pure Visual Cues 📈 Scalable Difficulty Dive in 🧵

wait for it..... wait for it......... now, there it is. Something fun is cooking!

@Songwei_Ge @jbhuang0604 @akanazawa @tomgoldsteincs @reveimage Congrats, Songwei!

Successfully defended my PhD thesis, "Visual Content Synthesis at Scale," today!! Huge thanks to my advisors @jbhuang0604 and David & committee @akanazawa @tomgoldsteincs and Shurva. ❤️ Super excited about the coming adventure at @reveimage! Slides: drive.google.com/file/d/1raqQlZ…

Attention sinks in LLMs are weird. There’s ~20% of heads that don’t seem to do anything. Do these heads matter? Turns out that if we get rid of them, benchmark scores don’t change.

Looking at the reviews in ICML, I am noticing more and more that some reviewers are assuming knowledge or rumors that may or may not exist in industry labs. This isn't great for open research

@Songwei_Ge @reveimage Congrats, Songwei!

Congrats to the team for the amazing results!! 🚀 Reve is truly incredible— I've had so much fun playing with it and creating visuals for different purposes. I'm also excited to share that I will be joining @reveimage this May and can't wait to be part of the cruise!!

My team at @GoogleDeepMind is hiring. If you are passionate about robust ML, the provenance of synthetic media, and the trustworthiness of data, consider applying: boards.greenhouse.io/deepmind/jobs/…

Ok, so I can finally talk about this! We spent the last year (actually a bit longer) training an LLM with recurrent depth at scale. The model has an internal latent space in which it can adaptively spend more compute to think longer. I think the tech report ...🐦‍⬛

New open source reasoning model! Huginn-3.5B reasons implicitly in latent space 🧠 Unlike O1 and R1, latent reasoning doesn’t need special chain-of-thought training data, and doesn't produce extra CoT tokens at test time. We trained on 800B tokens 👇

📢 We're hiring a Postdoctoral Associate to research 3D scene reconstruction, novel view synthesis, and inverse rendering. Join our team and contribute to cutting-edge projects in computer vision! 🔗 go.umd.edu/PostDoc2-2025

@NehaKalibhat @GoogleDeepMind Congratulations, Neha!

Life update: I recently graduated with a PhD and moved to New York to join @GoogleDeepMind! It’s been a journey of growth - overcoming rejections and imposter syndrome along the way. I leave grad life feeling humbled and grateful. On to the next chapter! 🚀

@DivyaKRaman1 @Dolby congratulations, divya!

Thrilled to announce that I have joined Dolby Laboratories @Dolby as a senior researcher, where I'll be working on generative AI!! Excited for this new chapter and looking forward to contributing to an amazing team!! 🚀

Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives Alex Hanson, Allen Tu, Geng Lin, @vasusingla71, Matthias Zwicker, @tomgoldsteincs tl;dr: SnugBox+AccuTile->precisely localize Gaussians; Soft+Hard Pruning arxiv.org/abs/2412.00578

Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives Contributions: 1. SnugBox: A precise algorithm for computing Gaussiantile bounding box intersections. 2. AccuTile: An extension of SnugBo for computing exact Gaussian-tile intersections. 3. Soft Pruning: An augmentation for pruning Gaussians during densification. 4. Hard Pruning: An augmentation for pruning Gaussians post-densification.