Status: this experiment has concluded and its sample code has been removed from the repository; the HQ cloud path it compared against now lives in the engine (see HQ Volumetric Clouds). The write-up is kept for its findings.
Companion to samples/hq_cloud_pass.ts. This sample tests one
hypothesis: the HQ pass's lighting is good, the HQ pass's plumbing is what
breaks — so put HQ lighting on the cheaper pass's volume and skip the plumbing.
The two starting points#
CloudPass (engine) | HQCloudPass (sample) | |
|---|---|---|
| Resolution | Full-res, one pass | Half-res march + full-res temporal reproject |
| Volume | Flat slab, ~200-unit march clamp | Spherical shell, horizon-distance march |
| Density | cd_sample_density straight | hq_density + domain warp + mip LOD |
| Phase | Henyey-Greenstein dual lobe | Cornette-Shanks + back lobe |
| Self-shadow | 2 straight taps | 5-tap cone + long tap |
| Scattering | Single-scatter + ambient | 3-octave Wrenninge multi-scatter, Beer-Powder |
| Status | Ships in crafty, looks good | WIP — three unresolved artifacts |
The HQ pass's three open issues — axis-aligned puff rows, reproject latency, distant shimmer — are all properties of the right-hand plumbing, not the lighting:
- Rows come from the long-range march resolving the tileable 64³ noise grid. The HQ pass added a domain warp + mip LOD to fight it; not fully solved.
- Latency comes from the temporal reprojection leaning on history.
- Shimmer comes from sub-pixel detail under the half-res jitter.
What this sample does#
MergedCloudPass (samples/hq_cloud_merged/merged_cloud_pass.ts) is the engine CloudPass
structurally — same four bind groups, same overlay + godray-fog plumbing, same
single full-resolution render pass — pointed at
clouds_hq_merged.wgsl. The shader keeps the left
column (flat slab, cd_sample_density, short march clamp, white-noise dither)
and grafts in the right column's lighting only:
dual_phase→ Cornette-Shanks forward + soft HG back lobecone_light_march→ sun-perpendicular 5-tap cone + long tap (replaces the 2-tap straight march)multi_scatter_octaves→ Wrenninge/Hillaire 3-octave energy-conserving sumpowder_term→ Beer-Powder sun-edge darkeningmultiScatterblends the octave term against single-scatter, so0reproduces the CloudPass look and1is full HQ.
Because there is no half-res, no reproject, and no de-tiling machinery, none of
the three artifacts can appear in their original form. The single remaining
lever is marchDistance (the slab clamp): short keeps the tiling hidden the way
CloudPass does; long shows more deck and is where tiling rows would re-emerge.
Lower the deck (MIN/MAX H sliders) to pull the whole scene into the clean
short-range regime CloudPass was tuned for.
How to read the result#
Open hq_cloud_test.html and this side by side at the same sun angle. Same lighting math; the merged pass should be free of the row banding and the pan latency. The cost: no temporal supersampling, so the white-noise dither is visible without TAA — leave the TAA toggle on (it averages the dither the way CloudPass relies on a fine perceptually-flat grain).
Open questions this experiment is meant to answer#
- At what deck height /
marchDistancedoes the flat-slab tiling start to read as rows again? (i.e. how much of the HQ look survives at full range.) - Is the cone self-shadow march stable at full-res single-pass, or does it need the temporal average to stop fireflies on the cone taps?
- Does single-pass multi-scatter cost more or less than half-res + reproject in practice, for equal visual quality?