The classical capture pipeline — structure from motion, dense reconstruction, meshing, texture baking — is a genuine achievement and it is being displaced for a large class of uses, fast.
Not because splats are better at everything. Because for the specific job of capturing a real space and letting somebody walk through it, they win on the axes that turned out to matter.
What actually changed
Time to first look. Mesh pipelines are batch jobs. You capture, you wait, you find out you missed a wall. Splat optimization converges fast enough that you can inspect a rough result while still standing in the room, and go re-shoot the corner you missed.
The materials you always lost. Meshing hates anything that is not an opaque diffuse surface. Glass, chrome, foliage, hair — all reconstruct badly, and the fixes are manual. Splats represent view-dependent appearance natively, so a window looks like a window from every angle instead of like a grey hole.
Failure looks different. A mesh that runs out of data produces confident geometry that is wrong: a smooth surface where nothing was observed. A splat reconstruction with insufficient coverage produces something visibly unresolved. Honest failure is worth a lot when a human has to decide whether to re-shoot.
A mesh fails by inventing a wall. A splat fails by looking unfinished. Only one of those tells you the truth about your capture.
What you give up
This is not a clean win, and pretending it is will get you burned.
- No surface. There is no mesh to collide against, measure, or hand to a CAD tool. If your downstream need is physics or metrology, you still need geometry, and you will end up extracting it — badly — from the splats.
- Editing is unsolved. Moving a chair in a mesh is a known operation. Moving a chair in a splat scene means segmenting a cloud of anisotropic blobs and hoping the view-dependent terms still make sense afterward.
- Size. A room is millions of splats. Streaming that to a headset over a wireless link is its own engineering problem, and it is the one currently eating most of our time on Splat Studio.
Where the pipeline is now
Phone capture, walk the room for two or three minutes, upload, optimize, view in the headset in the same session. Ninety seconds of optimization on a single consumer GPU gets us to something worth walking through; a few minutes more gets us to something worth showing someone.
The bottleneck is no longer reconstruction. It is delivery — level-of-detail selection and streaming order, so the headset can start rendering before the whole scene has arrived. Classic systems work, in a new costume.
The honest summary
If you need geometry, build a mesh. If you need somebody to feel like they are standing in a room you captured this morning, the splat pipeline is faster to run, more forgiving of hard materials, and more honest about what it does not know.
We expect the two to converge — splats for appearance, an extracted surface for everything that needs to touch physics. Nobody has made that seam invisible yet.