Kosta Derpanis (sabbatical in Munich 🇩🇪)

Lecture slides for my "Introduction to #ComputerVision" and "#DeepLearning in Computer Vision" courses. 🆕 Gaussian Splatting 🆕 Flow Matching 🆕 Sparse Autoencoders The included videos do not contain voiceovers yet, planned for a future revision.

3DV 2026 opens tomorrow! 🎉 Don't miss out on a great lineup of orals, posters, our first panel discussion during lunch, and the opening keynote by Christian Rupprecht. See the full schedule & booklet in the links below:⬇️

Today we release a new paper from Meta @AIatMeta: "Interpreting Physics in Video World Models," one of the first interpretability studies of video encoders. V-JEPA 2 shows rich, counterintuitive behaviors, including brain-like population codes and high-dimensional steering.

Yu et al., "MosaicMem: Hybrid Spatial Memory for Controllable Video World Models" A patch-based spatial memory that you raster into views + glues to make things work.

🤖 What would LMArena for robotics look like? Introducing RobotArena ∞ We turn real videos into simulated environments and evaluate robot policies at scale using VLM scoring + human preferences A scalable benchmark for robot generalists 🔗 robotarenainf.github.io Details 🧵👇

Long-LRM++ is our newest high-resolution feedforward scene reconstruction, to appear in CVPR 2026 (findings). With feature splatting, we got significantly more clarity — small font texts can now be clearly rendered in a few seconds!

Finally got some time to look at the popular Seoul World Model recently. What excites me most isn’t training giant WM; it’s anchoring generation to a map that retrieves from what we already see, and scaling up the pipeline. Very exciting! seoul-world-model.github.io

@ahmadsalimi_ Same happened to me in Sydney, Australia 😅

@CSProfKGD Escalators' orientation is the same I think haha. Repeatedly took the wrong escalator.

Note to self: People in the UK walk on the left … just like the cars. Explains why I keep bumping into people 😅

New episode of the Information Bottleneck! We talked with @StefanoErmon about why he thinks diffusion LLMs will replace autoregressive ones. Stefano co-invented DDIM, FlashAttention, DPO, and score-based diffusion models. He's a Stanford professor and now runs @Inception_AI, where they built Mercury II. We go deep but also cover the bigger picture - the startup journey, PhD vs industry, and where AI is heading. A few things that stuck with me: - He thinks of autoregressive models as typewriters and diffusion models as editors. One goes left to right. The other starts messy and refines. - Mercury II (their text difussion model) already beats the fastest autoregressive models on latency-critical stuff as voice agents, code suggestions, anything where you have a tight time budget. And it does it because diffusion generates tokens in parallel instead of one at a time. - We also got into whether AI will actually replace software engineers (his answer: no), PhD vs industry advice, and what it was like going from an ICML best paper to raising money.

Video might be the next intelligence substrate. Strikingly, video models are beginning to exhibit the same emergent reasoning behaviors first observed in LLMs—multi-path search, self-correction, and layer specialization. We demystify video reasoning and show it doesn’t happen frame-by-frame, but along diffusion steps. 🔗 wruisi.com/demystifying_v… 📄 huggingface.co/papers/2603.16… So, what's next? ;)

Ummm, so a typical conference 🤓

FIRST LOOK: Val Kilmer has been resurrected via AI to star in the new movie "As Deep as the Grave." Kilmer was cast in the movie in 2020, five years before his death. But he was too sick amid his throat cancer battle to ever make it to set. Now an AI version of the actor is appearing in the film, with the full blessing of his daughter, Mercedes: "He always looked at emerging technologies with optimism as a tool to expand the possibilities of storytelling. This spirit is something that we are all honoring within this specific film, of which he was an integral part.” “He was the actor I wanted to play this role,” says writer-director Coerte Voorhees. “It was very much designed around him. It drew on his Native American heritage and his ties to and love of the Southwest... His family kept saying how important they thought the movie was and that Val really wanted to be a part of this. He really thought it was important story that he wanted his name on. It was that support that gave me the confidence to say, okay let’s do this. Despite the fact some people might call it controversial, this is what Val wanted.” wp.me/pc8uak-1lH1PI

Introducing the new @stitchbygoogle, Google’s vibe design platform that transforms natural language into high-fidelity designs in one seamless flow. 🎨Create with a smarter design agent: Describe a new business concept or app vision and see it take shape on an AI-native canvas. ⚡️ Iterate quickly: Stitch screens together into interactive prototypes and manage your brand with a portable design system. 🎤 Collaborate with voice: Use hands-free voice interactions to update layouts and explore new variations in real-time. Try it now (Age 18+ only. Currently available in English and in countries where Gemini is supported.) → stitch.withgoogle.com

@AIPhDEthan That ship has sailed awards.acm.org/about/2024-tur…

Richard Sutton's RL work at UAlberta makes him a strong contender. That pipeline runs deep.

Another Canadian wins a Turing award 🇨🇦 I see a trend. Which Canadian is next? 🤔

An unsolicited guide to being a researcher: super instructive slides by @EugeneVinitsky emerge-lab.github.io/papers/an-unso… - different goals of a PhD student - how to be a good collaborator - how to keep up with literature - tracking your ideas & experiments - stress & productivity

Bring your research to life at #CVPR2026! The Call for Demos deadline is only 2 DAYS away. ⏳ Don't let your breakthrough stay trapped in a PDF. Show the community what your work looks like in action. Join us in Denver! 🚀 Submit here: cvpr.thecvf.com/Conferences/20…

@CSProfKGD I did a similar thing with copilot + opus/sonner. Wrote a blog post about it: itzikbs.com/blog/posts/202… Now updating anything is just another agent call away.

@ducha_aiki @CVPR

Sad, but important announcement -- one of our challenges -- Building3D lost their sponsor in the very last moment. So new prize fund is $14k: - $12k for S23DR 2026 - $2k for Building3D 2026. We apologise for this situation, and hope that you still decide to participate. #CVPR2026

I use Jekyll to build my academic site. Maintenance used to be painful. With Codex, it became painless.