3D Intelligence Report – September 10, 2026
Theme of the day: specialist work turning into a standard part you can call. A transformer renders caustics, fog and glass without a light-transport solver written per effect. Gaussian splats become an ordinary three.js object that raycasts and frustum-culls like a mesh. Wayve puts a staff-level engineer on sensor calibration, the plumbing that never used to come with a title. A sensor company raises 165 million dollars to do position and motion in silicon instead of in software. And the CVPR 2027 call lists Physical AI and world models as their own subject area, which is how a frontier quietly becomes a category.
Every link below was fetched and verified on September 10, 2026, the day this report went out.
RenderFormer-V2: Neural Rendering with Heterogeneous Scene Primitives
Chong Zeng, Yue Dong, Pieter Peers, Lvmin Zhang, Maneesh Agrawala
128K primitives per forward pass, 32 times version one. 3.92 seconds at 2048 pixels, against 53.8 for Blender Cycles.
RenderFormer treats global light transport as a sequence-to-sequence problem: triangles and their materials go in as tokens, pixels come out, and nothing in between is a hand-written ray tracer. Version two makes that idea usable on real scenes. A windowed attention pass along a Hilbert curve, plus a small set of attention sink tokens, drops the cost from quadratic in primitive count to something that fits 128K primitives in one pass instead of 4K. It also stops assuming every scene is made of opaque triangles: environment maps, participating media and displaced or textured surfaces are all first-class primitives now, and materials are encoded in a 9-dimensional learned appearance space rather than a fixed reflectance model, which is what lets it hold up on materials it never saw in training. Trained on roughly 10 million procedurally generated scenes, about 70 TB, referenced against Blender Cycles at 4,096 samples per pixel, 19 days on 32 A100s.
The number everyone will quote is the speed, and it is the least interesting part. 3.92 seconds a frame is not real time, and it is only fast next to a path tracer that was computing ground truth. What I keep turning over is that nobody wrote a caustic in there. Caustics, volumetric scattering, colour bleeding in a Cornell box, all of it falls out of one learned model that was shown enough physically referenced images, instead of out of a solver somebody hand-tuned per effect. That is the same trade we made when we stopped writing per-object segmentation rules. You give up the ability to reason about exactly why a pixel is that bright, and you get generality you could not afford before. The 9-dimensional material space is the piece I would watch: if appearance is a learned coordinate rather than a reflectance model, then editing a material becomes moving a point, and that is a very different tool than the one we have.
Staff Robotics Engineer, 3D Sensing
GBP 130,000 to 180,000 base (est.), plus equity
Calibration has always been the thing you do before the interesting work, usually by whoever is free, usually documented in one person's head. Wayve is putting a staff-level individual contributor on it with a roadmap and a mandate to mentor. That tells you where the error budget actually lives once your perception models are good: not in the network, in the six centimetres and four milliseconds between two sensors that each think they are right. The desirable list is worth reading twice, because it puts factor graphs and Gaussian splatting in the same paragraph. The classic estimation stack is not being retired in favour of learned 3D, it is being asked to hold hands with it. Second Wayve role in this pack since July, and deliberately so: the July one built the reconstruction, this one makes the input to it trustworthy.
CVPR 2027 call for papers
Not a grant. This is the publication venue, so the return is peer review, visibility and the credential rather than money.
A call for papers is a map of what a field has decided to take seriously. Physical AI and world models were a workshop topic two years ago and they are a listed subject area now, sitting next to the geometry track that has been there for decades. If you are holding a result and telling yourself it is not conference-shaped, this is the year the categories moved toward you. The OpenReview moderation window is the practical trap: two weeks for a profile without an institutional email means the real personal deadline for anyone outside a university is late October, not November 16.
three.js r186
Raycasting, frustum culling and spherical harmonics, on a splat
r186 lands a native Gaussian splat renderer and loader written in TSL, so the same code path runs on WebGPU and WebGL. GaussianSplatMesh gains raycasting, per-mesh bounding boxes and frustum culling, view-dependent colour through spherical harmonics, RGBA8 colour packing in storage buffers, and glTF import. Practically: a splat stops being a special viewer you bolt onto a page and becomes an object you can click, cull, transform and mix with meshes, lights and shadows in a scene you already have. The same release ships SunLight with cascaded shadow maps, voxel cone tracing for diffuse global illumination, and order-independent transparency.
Every splat viewer I have used so far has been its own island. It renders beautifully and then you cannot click anything, cannot occlude it properly, cannot ask which splat the cursor is over. Raycasting and frustum culling sound like housekeeping until you notice what they mean: the splat now answers the questions the rest of the scene graph has always been able to answer. That is the difference between a demo and a component. And it goes both ways, because splats inherit everything three.js already has, the loaders, the controls, the post-processing, the enormous pile of code people wrote for meshes. If you have been waiting for the moment splats stop needing a bespoke pipeline to appear on a web page, this is closer to it than any viewer release this year.
Lyte raised 165 million dollars in a Series C at a 1.6 billion dollar post-money valuation to scale LyteVision, a perception platform that measures position and motion in custom silicon instead of reconstructing them in software
LyteVision puts 4D coherent vision, high-resolution imaging and inertial measurement on one synchronised timeline in hardware, so the sensor reports position and motion directly rather than handing frames to a software stack that has to infer them. Led by Maverick Silicon, with Fidelity Management and Research, Atreides Management, Key1 Capital and Ora Global participating. Founded 2021 in Sunnyvale, 272 million dollars raised in total, already shipping into inspection, logistics and manufacturing, and it took a CES 2026 Best of Innovation award in robotics.
Look at what is being funded here. Not a better model, a better clock. The pitch is that if your camera, your depth sensing and your inertial unit share one synchronised timeline in hardware, a whole class of software work stops existing, which is exactly the work the Wayve role above is hiring someone to do by hand. Both bets are live at once and that is the interesting part: one company is paying a staff engineer to make imperfect sensors agree, another raised 165 million dollars arguing they should never have disagreed. For anyone doing 3D from commodity rigs, the useful question is not who wins, it is how much of your accuracy problem is genuinely algorithmic and how much is two devices that were never truly synchronised. I would want that number before I bought either story.
Get the next report in your inbox
Five verified finds, my take on each, one short email a day.
Five verified finds with my take, one short email a day.
