Real-time global illumination in the browser. Direct and indirect lighting are both resolved with ReSTIR (Reservoir-based Spatio-Temporal Importance Resampling) on WebGPU compute shaders, with a brute-force path tracer alongside as the ground-truth reference.
▶ Try it live — no install, just a browser with WebGPU (recent Chrome, Edge, or Safari). The page reports the reason if the API or an adapter is unavailable.
Cornell box, ReSTIR DI + GI. Colour bleeding from the red and green walls, soft shadows, and multi-bounce indirect light — all resolved in real time.
- ReSTIR DI —
Mlight samples per pixel resampled with RIS, then combined with the reprojected reservoir from the previous frame and with reservoirs from neighbouring pixels. Visibility is excluded from the RIS target function and applied once, on the surviving sample, at shading time. - ReSTIR GI — one indirect bounce per pixel produces a sample point carrying the radiance it reflects towards the visible point, reused temporally and spatially through the reconnection Jacobian, with a visibility ray guarding against light leaking between reused pixels.
- Denoised path tracer — one path sample per visible surface, accumulated temporally and filtered by the same à-trous chain. This is the equal-time comparison path: cheaper frames buy more independent samples, while the filter sees the same albedo-demodulated illumination representation.
- Reference path tracer — no resampling, no denoiser, one path per pixel per frame progressively averaged. Use it to check what ReSTIR converges towards.
- Every stage is an independent toggle — DI/GI, temporal reuse, spatial reuse, the à-trous filter. The contribution of each part is visible on its own.
- Analytic dielectrics — the glass scene adds a sphere and two clear cuboids with Fresnel reflection, refraction, total internal reflection, and subtle transmission tints.
Glass scene: refraction through analytic dielectrics, with the red wall inverted through the sphere.
The overlay reports resolution, frame time and how many frames have accumulated; the panel holds the per-stage toggles and sliders.
- Drag to orbit, scroll to dolly.
- The panel exposes RIS candidate count, spatial neighbour count and radius, GI bounce depth, accumulation window, resolution scale, and exposure.
- The room is a closed box; primary rays cull back-facing surfaces, so the near wall becomes a cutaway and the interior stays visible from any angle.
Six scenes are staged in the same unit-cube room, all primarily Lambertian so diffuse interreflection stays the core of the image:
| Scene | What it is for |
|---|---|
| Cornell box | The classic single ceiling emitter and two blocks |
| Glass sphere & cuboids | Three clear analytic dielectrics with reflected and refracted room detail |
| 30 lights | A grid of tinted emitters, where DI resampling has more to work with |
| Two rooms | A partition with one doorway: the near half is lit only through the opening and by bounce |
| Cove light | An upward-facing emitter behind a lip: the room below it is lit by the bounce off the ceiling |
| Pillars | Nine pillars under a broad emitter — overlapping penumbrae and contact regions |
Toolchain versions are pinned in mise.toml (mise
provisions them with mise install).
pnpm install
pnpm run devpnpm run lists the rest — build, test, test:e2e, lint and fix.
Vitest 5 runs source tests in the dom project (happy-dom) and root tooling
tests in the node project. Run either with pnpm run test --project dom or
pnpm run test --project node. Test APIs are imported from vitest explicitly.
Install Chromium once before running local E2Es: pnpm exec playwright install chromium.
CI provides Chromium through its Playwright container and retains failure traces
in the playwright-test-results artifact for seven days.
Two things the script names do not tell you:
E2E_PREVIEW=1 pnpm run test:e2ebuilds first and runs againstvite previewinstead of the dev server. That is what CI does, and the only way to exercise Worker asset serving and the hashed production bundle locally.- Development builds add an in-page renderer comparison to the stats overlay and
expose
window.__gifor scripted experiments — see manual comparison. WGSL compile errors reach the console with the shader name andline:column; without that they surface only as invalid-pipeline warnings at dispatch time.
Run the BDPT and GPU diagnostics E2Es on a local hardware GPU:
CI= E2E_PREVIEW=0 E2E_WEBGPU=1 pnpm run test:e2e --workers=1 'bdpt-.*\.spec\.js' gpu-diagnostics.spec.jsE2E_WEBGPU=1 enables Chromium's unsafe WebGPU flag and selects Metal on macOS.
The command uses the development server for harnesses that import source modules.
Tests skip without an adapter; the spatial-profile tests also need
timestamp-query. The hosted CI runner does not provide hardware GPU coverage.
main deploys to Cloudflare Workers; pull requests upload a preview version and
post its URL as a sticky comment. The conditions live in
.github/workflows/deploy.yml.
The repository secret CLOUDFLARE_API_TOKEN is an account-owned Cloudflare token
with the Individual Workers Editor permission scoped to the web-gi Worker only
(Specified Workers, Editor role). A Worker-scoped token cannot create a
Worker or other resources, so create anything new once with a broader token.
wrangler must stay at 4.156 or newer: older versions fail after the upload,
reading the account-level workers.dev subdomain with this token.
One rule in the workflow is worth stating because the workflow cannot explain itself: fork and Dependabot pull requests are deliberately skipped rather than broken. Neither can read the Cloudflare token, so a deploy step would fail on every such PR; skipping keeps that signal honest.
- Pipeline and architecture — the per-frame pass order, scene representation, and accumulation behaviour.
- Benchmark methodology — the equal-time comparison presets, how a verdict is decided, and why the tie threshold is provisional.
- Spatial reuse is bounded in world units — a unit mismatch that let reuse reach across the room, and the fix.
- The
M-weighted reuse combination is biased, and the visibility test that rejects GI candidates is not reflected in the 1/Z normalisation. The measured residual sits well under a percent of luminance in either direction — see Known bias. - ReSTIR reservoirs are restricted to diffuse vertices. Glass is evaluated as delta paths at shading time and never stored in a reservoir.
- Which renderer wins depends on the condition, so any claim about it has to name one — see the recorded runs.
- The à-trous filter ships as a texture-backed pass. A workgroup-tiled variant
exists behind
?atrous=tiled, but it measured slower on the device it was written for and needs more workgroup storage than WebGPU guarantees, so it stays an opt-in experiment rather than the default.
- Bitterli et al., Spatiotemporal reservoir resampling for real-time ray tracing with dynamic direct lighting, SIGGRAPH 2020
- Ouyang et al., ReSTIR GI: Path Resampling for Real-Time Path Tracing, HPG 2021
- Schied et al., Spatiotemporal Variance-Guided Filtering, HPG 2017
MIT