Pelican SDF benchmark

Skip to content

BENCHMARKSPEC v1.0.0

Pelican SDF benchmark

A reproducible test for model-generated WebGL2 signed-distance-field scenes. The model receives one blank canvas and must implement the renderer, scene, camera, and controls in a single response.

View results →<br>Methodology

InputHTMLCanvasElement

OutputSelf-contained ES module

EvaluationGPU timing + visual rubric

Host inputHTMLCanvasElement

Module exportcreatePelicanSdf(canvas)

RenderingWebGL2 · SDF · ray marching

Required inputpointer · wheel · pinch · keyboard

Reference viewport960 × 540

Source ceiling32 KiB

const run = createPelicanSdf(canvas)<br>run.start()<br>run.resize(width, height, dpr)<br>run.destroy()

INPUT1 blank canvas

LIMIT32 KiB source module

VIEWPORT960 × 540 reference size

CAPTURE12 views inspection set

METHOD

Methodology

The host provides no scene helpers. The submitted module owns all geometry, shaders, camera state, rendering, and interaction. The canonical prompt explicitly states that frame rate will be measured.

01Generate Fresh conversation, byte-exact prompt, first response only, no repair pass.

02Validate Compile and render checks, twelve camera views, close zoom, and required input paths.

03Measure GPU timer queries where available; synchronized wall-clock timing otherwise.

Full methodology ↗<br>Result schema ↗

CANONICAL INPUT

Build an interactive 3D signed-distance-field scene of a pelican riding a bicycle.

Host providesHTMLCanvasElement

Model providesEverything else

InteractionsOrbit · zoom · touch · keys

DisclosureFPS will be measured

RESULTS

Artifacts load on demand

The landing page imports no run data or WebGL modules. Open the results catalog to inspect artifacts, compare two runs, or access source and provenance records.

Open results →

The landing page is fully readable without JavaScript. Interactive benchmark runs are available in the results catalog.