3D Intelligence Report – October 29, 2026
Theme of the day: speed and trust between capture and deliverable. GPA-VGGT makes VGGT about three times faster on 500 frames without retraining. Waymo pays up to $263K base to evaluate automated HD maps. SuperSplat 3.7 becomes agent-drivable. SpAItial turns minutes of iPhone capture into a metric 3D world, and fills what you missed.
Every link below was fetched and verified on October 29, 2026, the day this report went out.
GPA-VGGT: Geometry-Preserving Acceleration of VGGT
Muyeol Sung, Jangwon Kim, Jongwoo Lim
500 frames in 159 s instead of 514 s
VGGT reconstructs cameras and geometry from a set of photos in one forward pass, but its global attention grows with the square of the number of frames, so hundreds of frames become slow. GPA-VGGT changes the attention without retraining. In the early layers, global attention becomes frame attention plus a shared set of cross-frame keys and values. In the later layers, keys and values are merged while every query is kept, so each pixel still gets its own output. On one RTX A6000, 200 frames take 33.65 s instead of 86.78 s (about 2.6x) and 500 frames take 158.63 s instead of 513.78 s (about 3.2x). Pose accuracy stays above 99% of VGGT: on CO3D with 200 frames, AUC@3 goes from 61.47 to 61.43. It runs the released VGGT checkpoint unchanged. Accepted at WACV 2027. Builds on the ledger's VGGT-Align (2608.15260, 2026-08-18).
A real survey set is rarely 20 photos, it is 300 or 500, and that is exactly where VGGT stopped being practical. This paper does not train anything new, it changes which frames talk to which inside the attention, and keeps the released weights. Going from 514 to 159 seconds on 500 frames is the difference between a coffee break and a quick check. My caution: the repo is three commits old and carries the VGGT License, so test it on your own sequences before you plan around it.
Senior Software Engineer, Mapping Data and Eval (Roadgraph)
$213K to $263K base, plus bonus and equity
Everyone wants to build the model that draws the map. This role is about the other half: proving the map is right, lane by lane, city by city. That is the same question a surveyor asks of any deliverable, only at a scale where you can no longer check by eye. If you know statistics and pipelines and you have been curious about mapping, note that they say no mapping experience is needed.
SuperSplat v3.7.0
Native WebMCP tools: an agent can edit the splat
Released 2026-10-26. Native WebMCP tools (PR #1081), a typed editor API built on serialized selection commands (PR #1078), invalid commands rejected instead of half-applied (PR #1082), copy and paste in the transform and colors panels (PR #1077). Fourth release in 24 days since the v3.0.0 WebGPU rewrite of 2026-09-08. MIT licence.
I covered SuperSplat in June and watched the WebGPU rewrite in September. This release changes who holds the mouse. With WebMCP tools and a typed command API, an agent can select, crop and clean a splat through the same commands you would click. The detail I like is that invalid commands are rejected, because an agent that half-applies an edit is worse than no agent. I would start with the boring jobs: removing floaters, cropping to the site boundary, the same cleanup on every capture.
SpAItial launches Echo-3 and an iPhone app that turns a few minutes of capture into a metric 3D world
SpAItial announced Echo-3 and the SpAItial app on 2026-10-27 ("your world, faithfully reconstructed"). A few minutes of iPhone capture gives a metric-scale reconstruction; Pro iPhones with LiDAR get live coverage guidance while you capture. iOS 18 or later, Android later. Subscription with the first month free, plus an enterprise API that accepts photos or video. Startup Fortune reports outputs as splats, point clouds and meshes; the blog itself does not list formats. The comparison against Polycam, KIRI and Scaniverse is the vendor's own. No accuracy figures are published.
The part I would underline is the gap filling. Echo-3 completes the parts of a room you missed, which is lovely for a walkthrough and dangerous for a measured drawing, because those surfaces were never observed. Metric scale from a phone is a real step. Before any survey or CAD use, I would scan one room I have already measured with a total station or a laser scanner and compare, and I would want to know which surfaces the model invented.
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.
