Ray-tracing renderer

This page describes the raytrace renderer selected by an ANARIFrame. See cameras, renderers, and frames for shared parameters and frame lifecycle.

const renderer = anariDevice.newRenderer('raytrace', {
  background: [0.012, 0.016, 0.04, 1],
  samplesPerPixel: 1,
  maxBounces: 1,
  progressive: true,
  shadows: true,
  resolutionScale: 0.5,
  minimumResolutionScale: 0.25,
  adaptiveResolution: true,
  targetFrameTimeMilliseconds: 33.3,
  temporalReprojection: true,
  shadowSamplesPerFrame: 1
});

frame.setParameter('renderer', renderer).commitParameters();

raytrace requires WebGPU. The ANARI adapter translates committed scene objects into descriptors for the shared RayTracingSceneRenderer in @luma.gl/anari. Its GPU compute graph derives world-space object bounds, Morton-sorts active object/instance leaves into an explicit retained permutation, builds and refits a graph-owned complete-binary TLAS, and traverses that hierarchy for nearest-hit rays and early-exit hard shadows. Transform-only animation gathers updated bounds through the retained permutation and refits without sorting; topology changes and periodic spatial refreshes rebuild the Morton order. A topology-only graph Morton-sorts each mesh's triangles into GPU-built BLASes, which transform-only updates reuse. It traces transformed analytic spheres and triangle meshes, including tessellated quads, cylinders, and cones; evaluates ambient, directional, point, and spot lights; and presents the result through a fullscreen pass. An rgba16float canvas or caller-owned framebuffer preserves HDR radiance; ordinary targets use the same configurable tone-mapping and exact sRGB transfer as forward rendering. Scalar metallic-roughness materials use GGX distribution, Smith visibility, Schlick Fresnel, and energy-balanced diffuse lighting. The trace pass uses exactly eight storage buffers, and every TLAS or BLAS construction pass remains within the default WebGPU CORE limit of eight storage buffers.

The default half-resolution internal target traces one quarter as many pixels as the output canvas. When adaptive quality is enabled, the renderer can reduce scale to 0.25, spread interleaved pixel coverage across frames, and rotate one shadowed direct light per frame. It approaches the configured frame budget using smoothed animation-frame intervals; no GPU timestamp feature is required. The fullscreen presentation pass upsamples the internal HDR result.

Progressive history is reprojected through previous camera matrices and stable ANARI instance/group/ surface identities. Depth and normal validation plus bounded neighborhood color clamping reject incompatible history; camera cuts, changed topology/materials, changed light counts, and target resizing invalidate it. GPU acceleration updates are encoded only for changed geometry or transforms, while camera-only and lighting-only frames reuse the retained TLAS. Transform-only frames use the retained-permutation gather/refit graph; topology changes and periodic refreshes use the full Morton build graph.

The Morton-sorted TLAS accelerates object and instance selection, while intersected meshes traverse GPU-built Morton-sorted triangle BLASes. Hardware ray tracing and SAH/Karras hierarchy topology are not implemented. Skeletal skinning, morph-target displacement, material textures, alpha blending/ transmission, and advanced PBR material extensions require raster rendering; deferred uses its forward fallback for unsupported materials. Diffuse secondary transport is opt-in through maxBounces (0–8). Each bounce samples a cosine-weighted hemisphere and gathers emission and existing direct lighting at the next hit, so matte surfaces receive color bleeding and emissive illumination. The scalar diffuse throughput in transportMode: 'diffuse' is base color × 0.96 × (1 − metallic). transportMode: 'pbr' additionally samples opaque GGX reflection; transmission is not sampled. Misses contribute no indirect radiance: background is a display color, not an environment light. Ambient lights and fog retain the existing per-hit approximations.

samplingSeed controls primary and diffuse paths. Pixel, frame, sample, and bounce indices separate samples; unchanged progressive frames converge through the existing bounded history. Seed and bounce-budget changes invalidate accumulation. With diffuse bounces enabled, geometry, placement, material and light changes reset history even when temporal reprojection is enabled: primary-hit metadata cannot validate indirect illumination. Camera movement resets history when temporal reprojection is disabled; otherwise it uses the existing primary-hit reprojection and camera-cut rejection. Use a stationary scene, full resolution, disabled adaptive resolution and temporal reprojection, and multiple progressive frames for comparison. Small emitters can converge slowly under the default BSDF-only emission gathering. Opt-in explicit emitter sampling, opaque GGX transport and spatial denoising are available; MIS, transmissive transport, advanced temporal denoising and volumes remain planned.

For the rationale behind the TLAS/BLAS split, Morton ordering, refit policy, megakernel execution, and temporal reconstruction policy, see ray-tracing technique background and tradeoffs.

Applications can also register custom renderer runtimes.