[DennisSmolek]

Adds examples/oidn-web-compare. It runs one path-traced frame through this library and through oidn-web, the denoiser three-gpu-pathtracer's OIDNDenoiser runs. It shows the result as a draggable split view or a |left − right| difference view, and includes a benchmark with timings and PSNR.

What it compares

• denoiser: runs on whichever runtime the page loads with (?runtime=).
• oidn-web 0.5:
  • its WGSL engine with the default tiles, which is what OIDNDenoiser uses;
  • its WGSL engine with wholeImage;
  • its WebNN engine.
• oidn-web 0.3.5: its last TensorFlow.js build, installed as oidn-web-tfjs, with default tiles and as a single tile.

oidn-web dropped TensorFlow.js in 0.4. OIDNDenoiser takes initUNetFromURL from the app, so the app chooses the oidn-web version. The tracer's own demo pins ^0.4.0.

Keeping the comparison fair

• Same weights. Every contender loads the same rt_hdr_calb_cnrm{_small,,_large}.tza.
• Same inputs. All contenders denoise one frozen frame with one rasterized albedo/normal G-buffer. Normals are written in both encodings: we map [-1, 1] ourselves, oidn-web expects OIDN's [0, 1].
• Same device and precision. One shared GPUDevice, fp16 for everyone (?precision=fp32 switches both). The TF.js build always runs fp32.
• One at a time. Each contender runs with nothing else on the GPU.
• Same clock. Wall time from the call until the output texture is written and the queue is idle.

Measured

The numbers are shown on the page from src/measured/*.json. Setup: Apple M-series (metal-3), Chrome 155, Eiffel scene, 8 spp, rt_hdr_calb_cnrm, fp16. Times are warm medians in ms over 7 runs; a range means the two runs differed.

size	ours WGSL	ours Auto (WebNN)	0.5 WGSL, default tiles	0.5 WGSL, whole image	0.5 WebNN	0.3.5 TF.js	0.3.5 TF.js, one tile
960×540	17.5	20.8	50–53	24.5–27.5	60–70	287	164–191
1280×720	32.0	29.9	79–95	50–65	95–106	598–624	263–279
1920×1080	89.5	50.5	251–316	115–134	220–231	1185–1241	fails (TF.js dispatch limit)

Quality is the same across contenders: 33.5–33.9 dB against a 2048-spp reference, and oidn-web's output matches ours to 52–57 dB.

Notes

• Reuses the Pipeline Lab scenes (../pipeline-lab/src/scenes.ts).
• Added to the CI typecheck list.
• Registered in examples.json with a shot but deployed: false. The production build passes (vite build), but I haven't verified it on Pages.
• Only measured on an Apple GPU so far.

Verification

• Checked in headless Chrome over CDP: live split, difference view, reference/PSNR and the full benchmark, with no WebGPU validation errors.
• tsc --noEmit, biome lint and vite build pass for the example.

🤖 Generated with Claude Code