RUST / WGPU · CLOUDFLARE R2
Load the time you need.
Seek a real STG-Lite model. Watch which chunks are requested, reused, and kept on the GPU.
Drag to orbit · scroll to zoom · Shift-drag or two fingers to pan. Arrow keys rotate, +/− zoom, Home resets.
Choose a scene, then load.
Address a time range
The server returns intersecting chunk descriptors. Each URL streams an immutable, checksummed R2 object and also supports HTTP byte ranges.
Open API responseNo request yet.
Spark & PlayCanvas: what we adopt · 200-second dataset & training plan / 渲染优化调研与长序列训练方案 →
Measured changes, including failures
CPU timings, GPU timestamps, image comparisons, source hashes, and dataset limits are published together. A lower point count or repeated short clip is never reported as a faster equivalent renderer or a longer dataset.
Same records, camera and time sequence. GPU mode culls and sorts on the GPU, then draws indirectly. No visible-ID upload; the count display arrives asynchronously. GPU total includes culling, sorting and rasterization.
GPU implementation, measured results and limitations →
CPU / GPU comparison results
Load a scene, then compare.
Published playback check: both scenes, 0 buffering events · Exact-record assembly A/B
Load a scene to measure frame intervals and buffering.
Load a scene to compare two 2-second idle windows.
Download benchmark report (JSON)LongVolcap / SelfCap · 4DGV / DeskGames · Data licenses & attribution