SwitchSplat brings switchable object states to 3D Gaussian Splatting, headed to SIGGRAPH Asia 2026
taiyasaki · x · 2026-10-20
Researchers from SFU, UBC, Wayve Labs and University of Toronto present SwitchSplat, the first 3DGS representation that models discrete object state changes, to appear at SIGGRAPH Asia 2026.
- Problem: existing 3DGS assumes a static world; time-conditioned methods like 4DGS conflate when a state was observed with what state the scene is in, so they can only replay seen sequences instead of composing arbitrary object configurations.
- Method: a base+slot decomposition—a shared static base captures the invariant scene while each slot independently holds the two binary states (on/off) of one interactable object; Gaussian-to-slot assignment is fully differentiable, so the model learns each state's geometry directly from casually captured video without manual annotation.
- Results: at inference users can freely combine slot states and render the scene from any viewpoint; the team released a casual-capture dataset and an interactive demo, with code coming soon.
More from Multimodal
- Midjourney sref 2081446411: jeweled crown portraits with lush film look — sergeantsref · 2026-10-20
- MOSS-Image-Base LoRA v14 released for realistic smartphone snapshot photos — AI_Characters · 2026-10-20
- MiniMax H3 ref2va struggles at video character replacement, user seeks workflows — alex13331 · 2026-10-20
- Qwen releases BlenderWorld benchmark for multimodal 3D creation agents — Qwen · 2026-10-20
- Reference-based image generation on 16GB VRAM: model and ComfyUI workflow advice — ApprehensiveRip4087 · 2026-10-20
- Stunning AI-generated gothic portrait wows the AI art crowd — Loo_Atreides · 2026-10-20