Skip to content

3D engine that AI agents drive

4.4 KB
Raw Blame

Verification

These figures are from the 0.3.x site and export work, measured around 2026-10-09 and not repeated since. Per-release checks from 0.4.0 on are in release-verification.md. Run npm test for the current test count.

The checks used Windows Chrome 154 through Playwright, an RTX 3060 through ANGLE D3D11 and CPU rendering through SwiftShader.

Check Result
Node tests Tests for geometry, bounds, scene budgets, timeline sampling and GLB loading passed.
GLB The independent Three.js GLTFLoader loaded all 574 WALL·E nodes. The assembled file contains 635,752 bytes.
Explosion The GPU pixel checksum returned to its original value after a 0 → 50 → 100 → 0 slider round trip.
Allocation reuse The shadow texture remained identical while the slider moved.
Layers and export Shell, interior and combined views worked. Exploded GLB translations matched the authored offsets.
Studio controls Finishes, keyboard orbit, reset, normals, auto orbit and downloads passed.
Recovery WebGL context loss and restoration passed without graphics errors.
Nature Daylight, wind, viewpoints, detail, water ripples, keyboard orbit and PNG download passed.
Film Out-of-order seeks reproduced ignition pixels. Playback, score loading, pause and mobile controls passed.
Layout Rendered studio, nature and film views were inspected at desktop and 320-pixel widths.
Reduced motion Nature starts paused. Film playback waits for a click.
Production Vite built all three pages. Its large-chunk warning applies to the optional Three.js benchmark.
Skills Both entrypoints passed frontmatter, length and route validation. GPU and CPU export commands produced usable files.

The studio review found panels covering the model and missing spaces around mobile line breaks. Camera framing and HTML whitespace fixed those problems. The subsequent screenshots show the repaired views. Material tags use flat shader varyings because interpolated tags previously produced stripes on the floor. The current studio enables four MSAA samples, FXAA and specular filtering.

Offline exports

Export Output Elapsed time Estimated render memory Observed Node RSS
GPU WALL·E image 1920 × 1080 PNG 1.68 s 119.4 MiB 124.0 MiB
CPU WALL·E image 640 × 360 PNG 4.02 s 22.7 MiB 124.5 MiB
GPU film, NVENC 1280 × 720, 32 s, 24 FPS 256.68 s 61.9 MiB 214.8 MiB
CPU film trial, libx264 640 × 360, 2 s, 12 FPS 7.33 s 22.3 MiB 157.5 MiB

The image runs use different resolutions. Their times do not establish a direct CPU/GPU speed comparison. Elapsed export time includes browser startup, PNG capture and encoding. All four runs reported zero WebGL errors. Both video runs held a maximum of one pending frame. The full GPU film sampled a maximum JavaScript heap of 6.9 MiB.

FFprobe verified 768 H.264 frames, 32.000 seconds, 24 FPS, BT.709 metadata and 48 kHz stereo AAC. The MP4 contains 13,759,581 bytes. Representative encoded frames were inspected in film-contact.png.

FFmpeg measured the encoded score at -21.0 dB mean volume and -4.5 dB peak volume. The score uses original synthesized tones and seeded noise. No human listening check was performed.

The JSON reports contain the exact measurements.

  • wall-e-gpu.json and wall-e-cpu.json record the image runs.
  • film-cpu-trial.json records the software rendering and encoding trial.
  • ../public/media/critical.mp4.json records the full hardware export.
  • benchmark.json records the comparison with Three.js.

The memory estimate covers authored targets, geometry and staging. It does not measure total driver VRAM or browser RSS. The sampled JavaScript heap can miss intermediate peaks. The guards reduce allocation risk without guaranteeing that every system allocation succeeds.

Limits

The benchmark measures CPU draw submission with identical flat shaders. Every result agreed on more than 99.99% of image channels. It does not measure GPU frame time or establish general Three.js superiority.

The film and nature world use stylized procedural assets. Blender-level realism remains unverified. Screen-reader behavior, 200% browser zoom and formal contrast compliance were not verified. The skills passed structural checks and real pipeline exercises. Automatic skill selection was not evaluated.