LTC Light Studio

Continuous light. Finite computation.
Polygon integration, fitted materials, and the cost of approximation.

Fitted GGX LTC

Compare the diffuse, GGX base and clearcoat integrals. Glass, film and ambient fill are shared.

THE IDEA IN THREE STEPSIntegrate a simpler shape. Keep the material.− / +

A light fills an angle.

Every surface point receives a continuum of incoming directions. Diffuse lighting integrates a cosine over that spherical polygon.

L_o = ∫_Ω L_e f_r cos θ dω

The fit changes coordinates.

A matrix warps a cosine lobe to approximate GGX. Roughness and view angle select a fit from the published 64² tables.

light polygon → M⁻¹ polygon

The boundary does the work.

The transformed polygon is evaluated along its edges. Monte Carlo samples the original direct-light model: differences expose fitting error and sampling noise.

surface integral → finite edge sum

Garage comparison: diffuse + GGX base + dielectric coat. Glass, film and ambient fill remain shared. Both direct surface estimators use unoccluded emitters. The optional numerical volume is shared.

Heitz et al. · SIGGRAPH 2016 ↗

From an integral to a boundary

Integrate the cosine. For an unoccluded polygon emitter, diffuse irradiance reduces to its spherical boundary.
Fit the glossy lobe. The garage reads published GGX LTC matrix/amplitude tables, transforms the polygon, and evaluates it deterministically.
Account for visibility. Unoccluded emitters keep the two estimators comparable. A retained experimental polygon-clipping path covers only aligned rectangular blockers.

E = \frac{L_{e}}{2} \sum_{i} θ_{i} \cdot (n \cdot ê_{i})

θᵢ = atan2(‖uᵢ × uᵢ₊₁‖, uᵢ · uᵢ₊₁)êᵢ = normalized edge cross productuᵢ: unit light directions, clipped above the horizon

This exact diffuse identity motivates LTC. The implemented GGX fit, edge approximation and horizon treatment are approximate; anisotropy is a heuristic extension. Runtime light integration is sample-free, not the entire image pipeline.

Heitz et al. · SIGGRAPH 2016 ↗

LIGHT / MATERIAL / CAMERA

Scene instrument

Lighting display

Integral Monte Carlo Compare

Experiment in this family

How does a fitted cosine lobe turn polygonal lights into a practical material model?

Animate light & scene

Extra fancy

Organic volume cube · raised-edge normals · flowing inlay

Backend

JS WASM GPU

Full material path

JS and WASM open the common diffuse studies at a smaller image budget. Returning to GPU restores this scene.

Material anatomy

GARAGE · GPU

Intermediate result

Floor roughness0.28

Normal detail1

Coat strength1

Material blending1

Texture frequency1

The live probe isolates the geometric floor’s base GGX fit. The image also includes mapped normals, procedural blends and a separate coat.

Atmosphere & motion

DETAILS +

Extra fancy adds a bounded organic volume. Turn it off to edit the room haze.

Medium extinctionOff Density structure

Scatter / extinction0.85

Noise scale1.00

Volume steps20

Ground flowOff

Animate atmosphereVolume onlyBronze light housing

Volume, surface detail & light animation ↗

Emitter geometry

LIVE

Light width 2.6 Light depth 1.7 Light height 2.9

Image & sampling

GPU BUDGET

MC samples 2 Exposure 1.15 Subpixel samples · GPU 2x

Render scale 3.00 Target FPS 60

Thin-lens camera

8 LENS SAMPLES

Aperture 0.000 Focus distance 5.20

Stylized ambient fill Horizon clip False color

Frame diagnostics

Surface hits 0

Display WebGPU

Resolution 960 × 600

Frame FPS 0

Producer CPU time 0 ms

Present submit 0 ms

Light work / lobe · estimate 0

Lighting closed form

Assumption direct + fill

Producer timing measures CPU work (GPU command submission on the GPU path), not GPU elapsed time. Garage light work estimates emitter samples or edge sets per lobe, before surface misses and varying material lobe counts.

Arithmetic throughput probe

A synthetic arithmetic kernel, separate from polygon lighting. CPU/WASM execution and GPU submission/readback have different measurement boundaries; this is not evidence of renderer speedup.

Selected WebGPU renderer

Kernel analytic

Kernel time pending

LTC direct surface lighting; numerical volume scattering and emission. No caustics or full global illumination. Implementation notes ↗

IMPLEMENTATION / COMPUTE

Three execution paths.

Feature support and speed are separate questions. These backends share a diffuse foundation, but do not implement the same complete material renderer.

Implemented in this project
Capability	JS	WASM	GPU
Diffuse polygon integral / MC	Yes	Yes	Yes
Fitted GGX, coat & material textures	—	—	Yes
Garage glass, aperture & atmosphere	—	—	Yes
Image execution	CPU	CPU	GPU

“Not applicable” means that feature has not been implemented in that backend here. JavaScript and WebAssembly can implement it; changing the execution language alone does not add it.

Why the scene changes

The CPU paths open a supported diffuse study, reduce resolution and disable GPU-only controls. The previous GPU scene and settings are kept for your return. The common scenes remain useful for comparing the same lighting integral.

Where the time goes

JS evaluates pixels in a main-thread loop. The compiled WASM kernel also runs on one CPU thread; it has no worker pool in this project. WebGPU dispatches many pixels concurrently. Resolution, light count and Monte Carlo samples multiply the work. WASM is not automatically faster than optimized JavaScript.

How much faster?

A recorded same-workload CPU audit measured 25.9 ms for JS and 4.5 ms for WASM: the moving trio at 192×100, split view, two MC samples. These are medians from this machine's Node/V8 run, excluding browser presentation—not a universal multiplier.

The live “Producer CPU time” includes actual CPU rendering for JS/WASM but only command submission for GPU. Dividing those numbers would not measure GPU speedup. The technical notes retain the complete measurements.

This panel stays open while you try the controls. Close it with the button above.

Backend audit & measurement method ↗

The material garage

A light fills an angle.

The fit changes coordinates.

The boundary does the work.