Hi all. I’m an architect in Gothenburg and I’ve spent the last year building Krill, a plugin that brings Gaussian splat scans into Rhino 8 as native block instances.
What it does: import a .ply scan, place it as a Krill block, and use it as site, context or interior. Layers, hide, isolate, undo and Grasshopper work as usual. Cut and crop volumes trim the scan live against your own Rhino geometry, and clipping planes work too. Point osnap snaps to splat centres, so you can measure on the scan. Inside a block edit you can delete stray splats non-destructively. Colour grading per placement, per block or globally. Export the trimmed, repositioned scan as .ply.
Krill renders with its own engine written for the Rhino viewport. Large scans build a level-of-detail pyramid on import; the clip below is about 600 million splats on an RTX 5060 Ti. Everything runs locally, nothing is uploaded.
Windows only, Rhino 8. AMD and NVIDIA tested. The package is on the Package Manager but locked behind an access code during the closed beta. If you want to try it, email krill@verkare.com
and I’ll send a code. Feedback from people with real scans and real projects is what I’m after.
No sorcery, but something I’ve dreamt about ever since I became a practising architect: to sample real spaces and draw and model architecture inside them.
To answer the question: none of the above. The splats aren’t mesh, Brep or SubD. What sits in the Rhino document is a block instance that Krill owns; the splats themselves are drawn by Krill’s own renderer straight into the Rhino viewport, with proper depth against your Rhino geometry, so solids occlude the scan and transparent objects tint it.
Corridor Crew’s video got a lot of people into this. If you ever want to try it, a phone capture through Polycam or Scaniverse is enough to get a .ply that loads. For a desktop workflow I use LichtFeld Studio (lichtfeld.io) - free and open source
I am curious cause I tried some photogrametry in the past to try to reverse engineer a boat hull, however, in the end, the final result made it harder than to just take measures to recreate the boat’s lines plan.
This Gaussian splats tech seens to be a bit lighter, in the matter of not creating a huge dense mesh to everything, so I thought it could make more sense to use it instead of trying photogrametry again.
Lighter, yes - there’s no dense mesh sitting in the document, and the level of detail keeps it moving whatever the size. But I don’t think it solves the hull. The splats sit near the surface, not on it, and for a lines plan that bit of noise is exactly the thing that hurts. You can snap to it and pull sections off as a starting point, but you’d still be fairing the curves by hand against real measurements. What it’s good for is having the boat there to model against, rather than to measure.
Worth adding: where splats have improved is on the capture side. Trained from photos alone they inherit the weaknesses of photogrammetry - the same drift and scale problems, since both start from the same structure-from-motion guess at where the camera was. Rigs that pair LiDAR with the cameras sidestep that: the LiDAR point cloud orients the images, so the splats are initialised from measured positions rather than inferred ones. It won’t make the hull fair itself, but it does mean the scan you’re modelling against is anchored in real dimensions instead of slowly wandering off over the length of the boat.
Close, yes - both are samples in space with no connectivity. But each splat has a size, an orientation, opacity, and a colour that shifts with the angle you look from, so instead of dots you get something that reads as surface. Krill only takes splat .ply though, no point clouds, no E57 or LAS.
No roadmap with dates on it. It’s a closed beta, the current build expires 30 November and a new one takes over before then. Price isn’t decided yet - the licence leaves room for both free and paid tiers. Beta testers will hear it first when it is. If you want to try it, mail krill@verkare.com and I’ll send you a code.
.ply isn’t Polycam’s - it’s the generic container most splat trainers write. Polycam is one of them, though .ply export sits behind their paid tier. There’s no recommended route as such, but Scaniverse is free, processes on the phone and exports .ply. On desktop I use LichtFeld Studio.
No renderer support - Krill draws in the Rhino viewport, so V-Ray won’t see it. V-Ray 7 reads splats on its own though, and .ply is the only thing its loader takes (docs), so export the trimmed .ply from Krill and bring that in. I haven’t tested that round trip myself.
And no mesh - Krill doesn’t turn splats into geometry. If you need one, run the photos through photogrammetry instead.
(Also, what software on the phone did u use? You mentioned POLYCAM-premium and SCANIVERSE, is the sample shown from one of these? Have you tried KIRI app?
Any reasonably modern phone or camera can work. I’ve tried Polycam, Scaniverse, KIRI and a few others over the years, though it’s been a while since I compared them directly. They all produce good results, and they’re constantly improving.
Personally I tend to prefer a more manual workflow: capture video, extract frames, run a COLMAP reconstruction, then train the splats in LichtFeld Studio. It gives me a bit more control over the process and tends to scale better to larger reconstructions.
The sample shown was not created from a phone app.
Hi Benjamin,
I am also very interested to give it a try. I’m using LiDAR and creating Gaussian splats in other software. I’d like to see this happening in Rhino. Could you send me your beta?
I own a revopoint scanner (pop4) that can also produce Splats, combined with the 3D pointcloud/mesh.
Later this week I will try if i can import a splat in Rhino and share the scan.