Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:

Thrilled to announce the @SpAItial_AI PhD Intern program. Work with a team of exceptional researchers. Build foundation models with lots of GPUs and publish the results at a top-tier conference. Incredibly happy to strengthen the connection between academia and frontier AI🤝

Sad news. Baldur's Gate and KotoR were so epic and they deserve so much credit for these pioneering games. Probably have to finish off Koveras and Irenicus one more time...

Congrats to @LukasHollein for his successful PhD defense! Lukas’s thesis is on 𝐆𝐞𝐧𝐞𝐫𝐚𝐭𝐢𝐧𝐠 𝟑𝐃 𝐖𝐨𝐫𝐥𝐝𝐬 𝐟𝐫𝐨𝐦 𝟐𝐃 𝐆𝐞𝐧𝐞𝐫𝐚𝐭𝐢𝐯𝐞 𝐌𝐨𝐝𝐞𝐥𝐬. His most impactful work so far, Text2Room (ICCV’23 Oral), was the first method to generate complete room-scale 3D scenes with textured geometry from text alone. The key idea is to lift pre-trained 2D text-to-image models into 3D: iteratively generate new views, estimate depth, and fuse everything into a seamless textured mesh, with a carefully designed viewpoint selection to complete the scene. This inspired many others on text-to-3D scene generation and led to a long line of follow-up works, including Ctrl-Room, LucidDreamer, WonderJourney, WorldGen, and many more. Today, the Text2Room paradigm of leveraging 2D priors has become the primary approach for both image and video diffusion models to generate 3D scenes. However, Lukas has done much more cutting-edge research. ViewDiff turns text-to-image models into 3D-consistent multi-view generators, WorldExplorer generates fully navigable 3D scenes from video diffusion models, and 3DGS-LM speeds up Gaussian Splatting optimization with Levenberg-Marquardt, and most recently World Reconstruction From Inconsistent Views. In summary, his PhD has helped define how we generate 3D worlds with generative AI. Lukas also keeps pushing the boundaries of 3D world generation. He’s now continuing his research at Apple and I'm sure we can expect more great research very soon! Check his great research: lukashoel.github.io Congrats Dr. Höllein, you will be deeply missed in the lab!

📢📢 Video Diffusion Avatars📢📢 Can video diffusion models (VDMs) replace 3D avatars? Video Diffusion Avatars (VDA) creates photo-realistic avatars from a single image, with full control over camera viewpoint, facial expressions and upper-body pose. Unlike prevailing 3D avatar methods, which commonly rely on 3DGS/NeRF, our VDM-rendering produces physically plausible dynamics of hair and earrings, as well as complex light interactions like shadows, reflections and refractions. Crucially, we show that common 3D representaion (e.g. 3DGS, NeRF) are not necessary for photorealistic and view-point controllable avatars. VDA is quite affordable for the research community, since training only requires 34 A100-GPU hours on public multi-view datasets, yet VDA generalizes far beyond them: to paintings, human-like creatures and full 360° orbits. 🌐 simongiebenhain.github.io/vi… ▶️ piped.video/llai-LGR6DI Great work by @SGiebenhain, @TobiasKirschst1, @liamschoneveld, @_davidedavoli_, @chenzhekl

Congrats to @Peter4AI for his successful PhD defense! Peter’s thesis is on 𝐆𝐞𝐧𝐞𝐫𝐚𝐭𝐢𝐯𝐞 𝐏𝐫𝐢𝐨𝐫𝐬 𝐟𝐨𝐫 𝐌𝐚𝐭𝐞𝐫𝐢𝐚𝐥 𝐄𝐬𝐭𝐢𝐦𝐚𝐭𝐢𝐨𝐧 𝐚𝐧𝐝 𝐋𝐢𝐠𝐡𝐭𝐢𝐧𝐠 𝐂𝐨𝐧𝐭𝐫𝐨𝐥. His arguably most impactful work, Intrinsic Image Diffusion, was among the first to bring diffusion models to material estimation. The key idea is that intrinsic decomposition is inherently ambiguous, so rather than regressing a single answer, it is formulated probabilistically and samples plausible decompositions, leveraging the strong image priors of diffusion models. This opened up a new line of diffusion-based inverse rendering work, including IntrinsicAnything, Marigold-IID, DiffusionRenderer, and many more. Today, generative priors are the standard approach for material estimation and relightable reconstruction. However, Peter has done much more cutting-edge research: LightIt brought explicit lighting control to diffusion models, IntrinsiX enabled direct PBR generation from text, and Intrinsic Image Fusion lifts these priors to high-quality multi-view 3D material reconstruction. In summary, his PhD has shaped how we think about materials and lighting in the age of generative AI. I’m also particularly proud that Peter is continuing this line of research as a Research Scientist at NVIDIA, e.g., his new work on Texture Space Material Diffusion and NeuMatEx. Also thanks to Ruediger Westermann and George Drettakis for being on the committee. Check out Peter's fantastic work: peter-kocsis.github.io Congrats Dr. Kocsis, it was an honor to have you in the lab!

We're building something very special ❤️ Can't share details yet, but it's a perfect moment to join us: we're hiring across world models, 3D diffusion, reconstruction, robot learning, graphics & ML infra.

📢OneCanvas (NeurIPS'26 Spotlight) Code Release📢 What if a 2D VLM could see the whole scene in one view? We reproject features from every frame into a single occlusion-free canvas to make a stock VLM reason in 3D! 🔗 github.com/baranowskibrt/one… 🌐 baranowskibrt.github.io/onec…

#1 customer question since 2017: “When can I send an avatar to a meeting for me?” Today, we’re launching @synthesiaIO Sessions - live Avatar conversations that help scale your time, so you can: - Train 1,000 sales reps in a day - Do 1:1’s with 2,000 employees at once - Interview 3,000 customers a week Try it free 👇