Something new is coming soon 👀

Finally, a look at SplatStroll. 🗺️ This project combines Spark 2.0 and dyno with CesiumJS and three.js to bring 3DGS to a global scale. Official support and a map-compatible demo are coming soon from @DuckbillStudio Learn more: studioduckbill.jp/3DGSLabo/S…

Spatial information for every device. @lileaLab demonstrates massive 40M+ splat scenes that run smoothly on smartphones and XR headsets like Vision Pro. Explore: works.lilea.net/spark/

@smallfly uses Spark LoD to turn real-world scans into "Dormant Memories." These are interactive, documented spaces exploring the intersection of memory and place. Experience it here: smallfly.com/dormant_memorie…

Spark 2.0 was built to make massive-scale 3D open and accessible on the web. It’s exciting to see builders already using the LoD system to stream high-fidelity worlds, like the 100M+ splat environments in STARSPEED. Take a look at how creators are shaping the future of 3D. 🧵👇

The final piece: Spark adapts OS-level Virtual Memory to the GPU. We manage a 16M splat physical pool that swaps 64K "pages" of data in and out from .RAD files using an LRU eviction strategy. This is how we render infinite, composite worlds on a mobile device. Demo: wlt-ai-cdn.art/spark-2.0/260…

Existing formats like .PLY and .SPZ aren't built for streaming. We developed .RAD: a spatially partitioned format that lets you see a 64K coarse version of a scene instantly while refining details as you move. Refinement happens in real-time, prioritized for your view. ⚡

Spark includes two "training-free" algorithms to build these hierarchies: Tiny-LoD for fast, on-demand use in the browser, and Bhatt-LoD for high-quality offline processing using statistical shape similarity. No reference images needed—just raw 3DGS data. 📁 Demo: wlt-ai-cdn.art/spark-2.0/260…

To keep the main loop smooth, Spark updates LoD updates asynchronously using Rust compiled to WebAssembly. We also use fixed foveated rendering to bias the splat budget toward the center of your view. More detail where you look, less where you don't. 🎯 Demo: wlt-ai-cdn.art/spark-2.0/260…

Performance Magic 🪄 Spark maintains a steady frame rate by computing "slices" through the LoD tree. By setting a budget (500K to 2.5M splats), we ensure a constant workload for the GPU regardless of scene size. You can trade detail for FPS in real-time to fit any device. Demo: wlt-ai-cdn.art/spark-2.0/260…