3D Intelligence Report – June 29, 2026
Five tracks ship today. Tool was nearly dropped (gsplat 1.5.3 dates to July 2025, FastGS to November 2025, the NVIDIA reconstruction toolbox to June 9), but a second pass on radiancefields.com surfaced COLMAP 4.1.0, released June 26 and squarely in window.
Every link below was fetched and verified on June 29, 2026, the day this report went out.
StructSplat: Generalizable 3D Gaussian Splatting from Uncalibrated Sparse Views
Zhao, Chen, Chen, Wang, Nie
28.05 PSNR on DL3DV, +5.67 dB over AnySplat
StructSplat is a feedforward 3D Gaussian Splatting model that takes a handful of ordinary photos with no camera poses and no intrinsics, no calibration of any kind, and predicts a 3D scene in a single forward pass. Every generalizable splatting method before it leaned on known poses or a separate pose estimator. StructSplat folds structure recovery into the network itself and still beats the current uncalibrated sparse-view leader, AnySplat, by 5.67 dB PSNR on DL3DV. It hit arxiv on June 26 stamped ECCV 2026, with the official training code already public.
Calibration has always been the quiet tax on photo-to-3D. For a century, measuring anything from images meant first measuring the camera: where it sat, where it pointed, what lens it carried. StructSplat is part of a wave that just deletes that step, inferring the geometry and the cameras together inside one network from a few uncalibrated shots. That is the smartphone-camera moment for reconstruction: the precondition everyone treated as unavoidable turns out to be learnable. The honest catch: the repo ships the training code but lists pretrained weights as coming soon, so today you can read the method and train it, not download a checkpoint and run it tonight. If you work with messy in-the-wild captures, this is the line of work to track, because the day the weights land is the day casual photo sets become 3D without a capture protocol.
Senior, ML Engineer – Neural Rendering
$177,300 – $234,000 (disclosed)
Read where the money is pointing, not just the title. Torc Robotics builds self-driving trucks, and this role is to take neural rendering, explicitly NeRF, 3D Gaussian Splatting and diffusion models, and build sensor simulation for camera, LiDAR and radar that trains and validates the trucks before they touch a real highway. That is the market signal worth more than the one opening: neural rendering is being hired not to make pretty novel views, but as safety-critical infrastructure for testing autonomy at scale. The salary is disclosed, $177K to $234K, which is a real anchor for anyone negotiating in this niche. If you go for it, lead with reconstruction work where the output had to be metrically faithful and feed a downstream system, not a benchmark you topped, because the whole job is whether the simulated world is trustworthy enough to bet a truck on.
SIGGRAPH Asia 2026 Technical Communications (Kuala Lumpur)
Peer-reviewed publication at a top-tier graphics and spatial-computing venue (no cash award); the short-paper track, lower barrier than the full Technical Papers track.
This is the most accessible serious-venue deadline open right now for this audience. Technical Communications is the short-form track: a focused result, a clever system, a working demo, not a year-long full paper, peer reviewed and published under the SIGGRAPH Asia name. For a practitioner sitting on one sharp 3D, splatting or spatial-computing finding, that is exactly the right size of bet. The deadline is July 27, twenty-eight days out, which is enough to write up something you have already built but not enough to start a project from scratch. The move this week is to look at what is already working on your machine and ask whether one piece of it is novel enough to stand alone, because the gap between a good internal result and a published one is often just the writeup.
COLMAP 4.1.0 4.1.0, with the Caspar GPU bundle-adjustment backend and 360 reconstruction
GPU bundle adjustment up to 100x faster than the Ceres backend
COLMAP, the structure-from-motion engine that sits under almost every photogrammetry and gaussian-splatting pipeline, shipped 4.1.0 on June 26. The headline is Caspar, a new GPU bundle-adjustment backend that the release notes put at one to two orders of magnitude faster than the CUDA Ceres backend on medium and large problems, the exact step that has always been the slow, CPU-bound bottleneck of incremental mapping. It also adds native 360 reconstruction with spherical (equirectangular) camera models, two new camera models (EUCM and division), advancing-front surface meshing, and gravity pose priors read from EXIF. Published release, downloadable binaries and source.
Here is a rabbit hole worth the fall. The solver COLMAP is replacing is named Ceres, after the asteroid that vanished behind the sun in 1801, and a young Gauss got it back by inventing least squares to fit its orbit from a few scattered sightings. That is bundle adjustment: take noisy observations of points seen from many viewpoints and solve for the geometry that best explains them. Two centuries later it is still the engine room of reconstruction, and still the slowest part. COLMAP 4.1.0 moves it onto the GPU with Caspar and reports one-to-two orders of magnitude speedups, which is the difference between a coffee break and a glance for a medium scene. The honest catch: that speedup is for medium-to-large problems on a capable CUDA GPU, so a tiny dataset on a laptop will not feel the same jump. If you build splats or meshes from COLMAP poses, the move is simple: rebuild a scene you know well on 4.1.0 and time the mapping step against your old run before you change anything in your pipeline.
At its June 22 keynote, SideFX made Gaussian Splats a first-class citizen in Houdini 22: a full toolset for rigging, animating, relighting, compositing and reconstructing splats, woven through the pipeline. Splatting is one of Houdini 22's two headline strategic initiatives, alongside AI.
Until now, Gaussian splatting lived in research repos and capture apps, a way to make a scene, not a native type in the tools studios actually ship with. Houdini 21 dipped a toe in with read-only visualization. Houdini 22 rebuilds it as a full citizen of the pipeline: you can rig a splat, relight it, simulate with it, composite it, generate it from renders. When a vendor as central to VFX and procedural 3D as SideFX bets one of its two headline initiatives on splats, that is the signal that gaussian splatting has crossed from a capture novelty into mainstream production infrastructure. The freshness note: this is a keynote sneak peek, the release is slated for later this summer, so it is direction confirmed, not software you install today.
The interesting moment is not the feature list, it is the word native. A technique becomes infrastructure the day it stops being a plugin you bolt on and becomes a type the whole tool understands, the way the layer became native to image editing in the nineties and quietly changed how everyone worked. Read-only viewing in Houdini 21 was a guest pass. Riggable, relightable, simulatable splats in 22 is a citizenship grant, and citizenship is what lets a representation pick up everything else the pipeline already knows how to do. For this audience the practical read is that splats are about to inherit the full weight of a production toolchain, which is the difference between a pretty capture and an asset you can actually work. Worth watching whether the other big DCC and game engines follow, because once a representation is first-class in one major tool, being absent from the rest starts to look like a gap.
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.
