No more slow models.
WhiskerSplat — fast, efficient, adaptable, powerful, robust.
A feed-forward model that reconstructs metric-scale 3D from as few as three unposed views in ~0.9 s — no per-scene optimization. State of the art on CT-Scenes-Hard, ~40× faster than optimization baselines. A fictional demo of the prism-glass theme.
About 40× faster than optimization baselines, on a single GPU.
PSNR on CT-Scenes-Hard — +0.99 dB over the strongest baseline.
One forward pass.
No test-time optimization: the network predicts the 3D field directly from the input views. Three properties carry the result — all numbers fictional.
Feed-forward in a single pass — ~0.9 s per scene, with no per-scene fitting.
Property 01 · fictional
~40× less compute than optimization baselines; runs on one consumer GPU.
Property 02 · fictional
Holds up from 3 to 9 views, indoor and outdoor, with no retraining.
Property 03 · fictional
What it does.
Three capabilities, measured on CT-Scenes-Hard. The links open the matching demos.
From 3 unposed RGB views to a metric-scale 3D field — drag the live reconstruction in the project page.
Open project page0.9 s / scene feed-forward inference; watch the training and ablations on the dashboard.
See the dashboardA single forward pass replaces minutes of per-scene fitting — the headline behind "no more slow models".
Read the posterEverything here is fictional.
WhiskerSplat, its authors, dataset, venue, and every number are invented to demonstrate the prism-glass theme.
WhiskerSplat is not a real model. People (Kitty Park, Ya-ong Kim, Calico Lee, Mittens Choi), the CatTower Vision Lab, the CT-Scenes dataset, the PURRCV 2026 venue, and all metrics are invented for demonstration. Nothing here is a claim of real performance, and no affiliation with any real entity is implied.