What guides the raster image?

Three routes, different contracts

The old modes are useful artistic previews, but a display-mapped tracer color is not an irradiance map: multiplying it by albedo/facing again is an appearance adjustment. The old branch also reuses the shared tracer state; it is not a separate, tightly specified surface-only estimator. Enhanced disables that old guided branch while active.

Enhanced: surfaces → transmission → scattering

1. Raster geometry and materials: draw at full display resolution. Preserve raster shading where the guide has no matching receiver.
2. Vacuum surface guide: progressively trace surface appearance. Its default scale is 0.5 in each dimension, subject to a 786,432-pixel budget with aspect ratio preserved.
3. Photon scattering film: run the selected photon-primitive family in a separate integrator. This CPU film is capped at 65,536 pixels and uploaded as linear RGB. It does not add a second copy of surface shading.
4. Composite: compute camera-to-receiver optical depth through supported homogeneous boxes/spheres. Apply transmission to the surface layer; add the already attenuated scattering film; apply exposure, tone mapping and gamma once.

C = exp(−τ) · mix(raster surface, accepted guide, blend) + scattering

At guide blend 1, accepted pixels use the traced surface estimate; full-resolution raster visibility still governs admission. This cannot recover fine shading detail absent from the low-resolution guide. The medium film has a separate depth-aware screen filter.

Camera, scene and transport-setting changes invalidate accumulation. There is no motion-reprojected temporal history in this compositor. Display blend, exposure and screen blur can change without invalidating the transport film.

Photon execution is another independent choice

• Independent: sample primitives independently for camera samples and accumulate over passes.
• Shared batch: gather the same primitive batch across the image. Normalize by attempted draws, not by successful intersections.
• Raster where implemented: use conservative raster gathering for supported records, including UV plane, beam and analytical hourglass. Unsupported families fall back to shared gathering; the status reports the actual execution.
• Fast hourglass / mesh sheets: triangulate finite refracted sheets and rasterize their contribution. Angular resolution, filtering and grazing regularization make this a biased structure preview. Here it supplies refracted caustics; its ordinary-fog and surface companion layers are disabled.

A medium-method MIS choice combines that family's supported strategies. It does not mean that the surface guide and raster image are MIS-weighted, or that every family has a raster implementation. A selected caustic-only method can leave ordinary direct fog absent.

What remains approximate?

This is a pinhole appearance preview. The vacuum surface guide plus camera-leg extinction does not reproduce extinction on secondary reflected/refracted surface paths, complete surface–medium exchange, depth of field, or a separate directly visible volume-emission layer.

The current GPU surface guide lacks arbitrary triangle-mesh intersections. Depth rejection prevents painting a traced background onto an unmatched foreground mesh, but cannot supply missing mesh shadows, reflections or indirect transport. The mesh-sheet caustic path has its own triangle intersections.

Heterogeneous, simulated, procedural and density-node media retain the old raster volume preview, with the vacuum guide disabled. Overlapping homogeneous optical depths are added in the composite; this can differ from the native medium-selection convention. More samples reduce noise but do not remove compositing and filtering bias.

Existing native captures

Retained captures from 13 September, not newly rendered benchmarks. The native comparison page records sizes and synchronized stage timings. Enhanced is not guaranteed faster than tracing: it still pays for a trace plus rasterization and compositing.

Environment LT: coverage does not imply efficiency

The new sampler makes environment light available to pure LT. It samples a uniform sky direction and a uniform entry disk around the scene, then traces inward. This is a real source sampler, not just a background painted behind geometry.

Many emitted rays miss useful geometry, hit occluded regions, or produce little contribution to the current camera. A loose scene bound wastes disk samples, while uniform directions spend samples on dim parts of a gradient sky. Eye-side sampling starts at a visible receiver, so it can spend much more of its effort on the image's relevant transport.

Thus ordinary broad environment illumination is often a poor use of this LT sampler. That is not a universal failure of environment light tracing: concentrated specular transport can favor light-side paths. Directional importance sampling and tighter bounds can improve efficiency; environment emission alone is not either of those optimizations.

Our radius test preserved mean brightness while larger-radius runs had higher receiver RMSE, consistent with wasted launch area. It was a normalization regression, not an equal-time benchmark against an optimized environment sampler. See the measured results and limitations.