Reality capture
Splat training studio
Train a 3D Gaussian Splat in the browser — from photos, while you watch.
What this is
A working 3D Gaussian Splatting optimiser: eight ray-traced views go in, anisotropic Gaussians come out, and every number on screen is measured during the run.
This is the pipeline LichtFeld Studio, Inria's original 3DGS and gsplat all implement, shrunk until a single CPU core can carry it. The optimiser is handed eight 64×64 images and their camera poses — nothing else, no geometry — and fits a cloud of anisotropic Gaussians to them: Σ = R S Sᵀ Rᵀ projected to a screen-space conic, depth-sorted alpha compositing, an analytic backward pass down to log-scale and quaternion, Adam, and adaptive density control. Training runs in a Web Worker so the viewer never drops a frame. The viewer is a real splat renderer — the covariance is projected in the vertex shader and the Gaussian evaluated per fragment — which is why turning the depth sort off is a control rather than a bug.
Things worth trying
- Press Train and watch the target photo and the current fit converge side by side
- Turn the depth sort off — sorted compositing is the algorithm, not an optimisation
- Switch to ellipsoids to see the actual Gaussians rather than their footprints
- Set densification to Aggressive and watch the export size climb faster than the PSNR
- Reset onto the room corner: flat painted walls give the optimiser almost nothing to work with
Bring us the product you were told could not be configured.
Tell us the option count, the CAD you already have and the frame budget you need to hit. An engineer reads it and replies with a straight answer — including when the answer is that good photography would serve you better.
We reply to every enquiry within one business day, in your timezone — with a frame budget, not a brochure.