Gaussian splatting has become remarkably forgiving about how a scene is captured, but it still expects the scene itself to hold still. Open a cabinet halfway through a walkthrough, switch on a lamp or pull out a chair, and a standard reconstruction either blends the two versions into ghosted geometry or discards one of them as a distractor. SwitchSplat, a SIGGRAPH Asia 2026 paper from Simon Fraser University, the University of British Columbia, Wayve Labs and the University of Toronto, keeps both.
SwitchSplat divides a scene into a shared static base, which holds everything that never changes during the capture, and a set of slots, one for each object the person filming interacted with. Each slot stores two versions of its object, off and on, so a closed door and an open door can both be kept in the same model. Every splat carries a small learned score that decides whether it belongs to the base or to one state of one slot. The authors describe the result as the first Gaussian splatting representation to address this problem.
That assignment is fully differentiable, so the model works out which splats belong to which state directly from the footage, with no object masks or manual labels. Each slot also learns a time window marking when its object was in the on state. An off the shelf AI vision model reads a sampled run of frames and proposes rough start and end points, and training refines them from there. The implementation builds on NVIDIA's 3DGRUT codebase, which reached its first stable open source release in April 2025.
The authors set this against 4D splatting methods, which tie the scene to a timestamp. In their framing, those methods conflate when a state was observed with what state the scene is in, so they can replay a recording but cannot produce a combination of states that never existed at the same moment. SwitchSplat treats every slot as an independent switch. The project page puts it simply, "Four parts make 16 combinations. The video shows only 5."
The team captured a new dataset of 12 scenes with a consumer Insta360 X4, walking through kitchens, flats, a balcony and around a car while opening and closing things along the way. Across the ten standard scenes, SwitchSplat averages 32.65 dB PSNR, against 30.55 dB for the strongest baseline, the dynamic 4DGS method, with the dynamic DeGauss and the static FullCircle further behind. On two scenes recaptured in object combinations that never appeared during training, it also scores higher in PSNR than both dynamic methods, even though those were allowed to pick their best matching moment in time for every test image.
SwitchSplat is limited to two states per object and assumes a single operator capturing with one free hand. Continuous or multi state objects, and discovering how many interactions took place without the AI model to seed the count, are left to future work.
The capture dataset is on Hugging Face, and an interactive viewer runs the scenes in the browser with every object state switchable. The code is available now on GitHub under the Apache 2.0 license.
