3D Intelligence Report – June 16, 2026
The Gaussian splat is quietly graduating from a research trick into a portable, first-class data type, and today's five signals land all along that path. One photo now becomes a splat inside a node graph anyone can run on their own machine. A geospatial pipeline drops the mesh step so point clouds and splats travel as themselves, straight to the viewer. A fresh paper pulls light and material back out of city video, so a scene can be relit instead of frozen on the day it was captured. A frontier lab is paying top of band for the metric localization that makes any of this trustworthy. And the single best research fellowship for this kind of work is open until September. The cost of making, moving, and lighting a splat just dropped, all at once.
Every link below was fetched and verified on June 16, 2026, the day this report went out.
BRDFusion: Physics Meets Generation for Urban Scene Inverse Rendering
Yi-Ruei Liu, Jie-Ying Lee, Zheng-Hui Huang, Yu-Lun Liu, Chih-Hao Lin
Relight a city from video
Inverse rendering pulls the separate ingredients of a scene, the materials and the lighting, back out of ordinary video, so you can relight it instead of being stuck with how it looked on capture day. BRDFusion fuses a physically based renderer (which respects how light behaves but breaks on messy real footage) with a generative model (which looks convincing but drifts), using each to cover the other's failure. It targets urban scenes from video and does novel-view relighting, night simulation, and dynamic object insertion. The code, the pretrained checkpoints, and the dataset are all released.
Inverse rendering is the unglamorous half of 3D that decides whether a capture is reusable. If you only have a splat, you have one frozen lighting condition: the street exactly as it was that one afternoon. Recover the materials and the light as separate things and you can relight it, drop it into a night scene, insert a new object that casts the right shadow. The honest tension is the one the paper names out loud: pure physics breaks on real messy video, pure generation hallucinates. Fusing them is the bet worth watching. Run it on your own street-level capture and check the shadows first, because shadows are where inverse rendering usually lies.
Senior Computer Vision Engineer (Localization)
$189K-$255K base + bonus + equity (disclosed)
Read the requirements, not the title. They want Structure from Motion, feature extraction, and Gaussian splatting rendering fused into a production visual positioning system that tells a device where it is to the centimeter. That is the quiet center of spatial AI: not generating worlds, but knowing precisely where you stand inside one. The comp is disclosed and serious because metric localization is hard, unglamorous, and load-bearing for everything built on top. If you apply, lead with a pipeline that held up outdoors, across seasons and changing light, not a clean indoor demo.
Marie Sklodowska-Curie Postdoctoral Fellowships 2026
EUR 399M call, around 1600 fellowships, 1 to 3 years, European and Global strands
This is the single best-fit funding door open right now for anyone doing 3D or spatial AI inside a research setting. It funds the researcher, not a consortium, and it pays you to move: to another country, often into or out of industry, to pick up skills you do not have yet. The mistake every applicant makes is treating the writing as the hard part. It is not. Most of the score rides on the supervisor and host match, and lining that up takes weeks. September 9 sounds far away. Finding and convincing the right host is the real bottleneck, so start that conversation this month, not in August.
ComfyUI v0.23.0
One image to a 3D splat
ComfyUI v0.23.0 ships day-zero native support for TripoSplat, Tripo's open-source (MIT) image-to-Gaussian-splat model, with a ready-made workflow template. From a single image you get a 3D Gaussian asset, with Gaussians allocated by where the image has detail and a control to dial detail up or down. The model weights live on Hugging Face and the inference code on GitHub, so it runs locally and wires into any custom node graph. This is the same TripoSplat we flagged as a watch item on June 15, now one click away inside the most used node editor in open-source AI.
The thing that matters here is not ComfyUI itself, it is the distance from a single photo to a usable 3D asset collapsing into one node graph anyone can run on their own machine. TripoSplat predicts the splat, ComfyUI turns it into a building block you can chain into the rest of your pipeline. The honest limit: single-image reconstruction is a confident guess about the side you cannot see, so treat the output as a fast draft, a previz asset or an AR placeholder, not survey-grade geometry. But as a way to go from a concept image to something you can drop into a scene this afternoon, it is hard to beat. Feed it a product shot and an oblique building photo back to back, and watch where the guess holds and where it invents.
Cesium ion reconstruction now runs point cloud and Gaussian splat jobs with no mesh step
For years a point cloud or a Gaussian splat had to be turned into a mesh before it could stream as 3D Tiles into a globe or a digital twin. Cesium ion now runs point-cloud-only and splat-only reconstruction jobs with no mesh step, so the raw geometry travels straight to the viewer. That removes a lossy conversion for survey and reality-capture data, and it signals that splats are turning into a native delivery format in the geospatial stack, not just a research output.
For years the unspoken rule was that everything had to become a mesh to be real in a geospatial pipeline. A point cloud or a splat got converted just to stream into 3D Tiles, and you ate the loss. Cesium ion dropping that mandatory step says the raw geometry is now trusted to travel on its own, all the way to the viewer. For anyone building a digital twin on survey data, that is one fewer place where the scan quietly degrades, and a clear signal that splats are being treated as a delivery format, not a research demo. The thing to watch next is whether the editors and analysis tools start expecting splats as input too.
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.
