Rust · wgpu · WGSL · measured in-browser
A small renderer experiment for asking one concrete question: what should a cross-platform viewer do with a time-varying Gaussian scene before reconstruction, compression, and production constraints are known?
Scope: procedural spacetime splats, not a learned reconstruction and not Peripheral’s implementation.
Use a current Chrome, Edge, or Safari build with hardware acceleration. The technical notes remain available below.
What this is—and is not
Every Gaussian here has spatial attributes and a time-dependent motion rule. Rust evaluates the scene at time t, culls and sorts visible splats, then wgpu submits instanced quads to a WGSL shader.
This proves the viewer and profiling path. It does not learn geometry or motion from video. A full 4DGS system additionally needs calibrated multi-view inputs, training or feed-forward reconstruction, a defined temporal representation, validation, and an asset pipeline.
Frame pipeline
Sample each splat’s position and opacity at the requested timestamp.
Use the same view/projection contract for culling, sorting, and rendering.
Avoid sorting and uploading splats that cannot contribute to this frame.
Back-to-front order makes standard alpha composition predictable.
Only visible GPU records cross the CPU→GPU boundary.
Instanced billboards evaluate elliptical Gaussian falloff per fragment.
Performance ledger
Live counters above are device-specific. Entries below distinguish implemented measurements from planned experiments.
| Status | Change | Why | Measurement / gate |
|---|---|---|---|
| Implemented | Frustum rejection before sort | Sorting invisible splats is pure CPU and upload waste. | Compare source vs visible count; prepare and sort ms remain exposed. |
| Implemented | Compact visible upload | The initial baseline rebuilt and uploaded every source record. | Upload/frame is shown in bytes; no hidden “optimized” claim. |
| Implemented | Reusable staging allocation | A new compact vector every frame is avoidable allocator churn. | The allocation was removed, but repeated wall-clock profiles did not establish a speedup; no performance win is claimed. |
| Baseline | CPU unstable depth sort | Simple, deterministic architecture establishes a correctness baseline. | Keep until sort + upload exceeds the frame budget on representative assets. |
| Planned | GPU preprocess + radix sort | Removes per-frame read/transform/sort pressure from the CPU at high splat counts. | Adopt only after an apples-to-apples trace shows lower frame time, not just higher complexity. |
| Planned | Compressed resident attributes | Large dynamic scenes are usually bandwidth- and storage-sensitive. | Track decoded bytes, visible quality, startup latency, and GPU memory together. |
| Planned | Screen-space covariance / EWA | The current billboard ellipse is a controlled approximation, not the full projected 3D covariance. | Validate projected footprint and aliasing before calling the renderer 3DGS-complete. |
Evidence boundary
Their public roles mention multi-view geometry, feed-forward temporal models, NeRFs, and 3D/4D Gaussian Splatting. Their Viewer role separately asks for a shared cross-platform rendering engine and spatial/video streaming.
No public source establishes that their live viewer uses one specific Gaussian variant, codec, sorting strategy, or renderer. CUDA and Metal are examples in the posting, not a published architecture.
It gives one typed rendering codebase over WebGPU in browsers and native graphics backends, which directly tests the cross-platform boundary in the Viewer role.