RT shadow denoiser everywhere: projectors (à-trous/SVGF) + inline-traced shadows; no TAA/FSR dependency #666

New issue

Closed

opened 2026-10-11 05:15:22 +00:00 by sky_line · 2 comments

sky_line commented

2026-10-11 05:15:22 +00:00

Owner

Owner direction: the RT shadow-mask denoiser (#85) is great and should be used everywhere. It removes the need for TAA/FSR to clean up shadow noise; both have their own problems and will be revisited.

Today

Ray-traced projector shadows go through the mask (RenderVKRTShadowMask) and get part of the denoiser:

per-light temporal accumulation (by light id)
variance clamp and fast reset on moving casters / disocclusion
a disc penumbra blur

They do not get:

the à-trous / SVGF spatial filter (RenderVKRTShadowAtrous: "Projector shadows keep the disc blur")
the SVGF change detection (RenderVKRTShadowSvgf, sun/moon only)
possibly the full anti-lag/reactive path (verify)

Inline-traced RT shadows (transparent surfaces, particles, the forward ray-query lit pass when the mask is off) get no denoising at all.

Goal

One denoiser path for every RT shadow signal: sun/moon, every RT projector slot, and anything else that traces shadows. That means temporal (adaptive history, clamp, anti-lag, SVGF change detection, reactive), à-trous spatial with variance, and the depth-aware upsample. Reuse the shared helpers clean_denoise_common.glsl / clean_rt_moving_caster.glsl from #657.

Projectors: per-slot variance plus à-trous iterations. Cost scales with the active RT projector count (capped); measure it and add sub-marks.
Alpha / particles traced inline: find a way into the denoised path, or document why not and the best alternative. Options: an alpha shadow-mask layer, or reuse of the opaque mask where depth matches.
Shadow output must stay unchanged where already converged. Keep A/B settings, no new per-frame allocations, VRAM within the existing accounting.
Without TAA/FSR (FXAA/SMAA/none), RT shadows must be as clean as with them. The denoiser must not depend on TAA's jitter/resolve; check the R2 noise phase and the reactive path.

Owner direction: the RT shadow-mask denoiser (#85) is great and should be used **everywhere**. It removes the need for TAA/FSR to clean up shadow noise; both have their own problems and will be revisited. ## Today Ray-traced projector shadows go through the mask (`RenderVKRTShadowMask`) and get part of the denoiser: - per-light temporal accumulation (by light id) - variance clamp and fast reset on moving casters / disocclusion - a disc penumbra blur They do **not** get: - the **à-trous / SVGF spatial filter** (`RenderVKRTShadowAtrous`: *"Projector shadows keep the disc blur"*) - the **SVGF change detection** (`RenderVKRTShadowSvgf`, sun/moon only) - possibly the full anti-lag/reactive path (verify) **Inline-traced** RT shadows (transparent surfaces, particles, the forward ray-query lit pass when the mask is off) get no denoising at all. ## Goal One denoiser path for every RT shadow signal: sun/moon, every RT projector slot, and anything else that traces shadows. That means temporal (adaptive history, clamp, anti-lag, SVGF change detection, reactive), à-trous spatial with variance, and the depth-aware upsample. Reuse the shared helpers `clean_denoise_common.glsl` / `clean_rt_moving_caster.glsl` from #657. - **Projectors:** per-slot variance plus à-trous iterations. Cost scales with the active RT projector count (capped); measure it and add sub-marks. - **Alpha / particles traced inline:** find a way into the denoised path, or document why not and the best alternative. Options: an alpha shadow-mask layer, or reuse of the opaque mask where depth matches. - Shadow output must stay unchanged where already converged. Keep A/B settings, no new per-frame allocations, VRAM within the existing accounting. - **Without TAA/FSR** (FXAA/SMAA/none), RT shadows must be as clean as with them. The denoiser must not depend on TAA's jitter/resolve; check the R2 noise phase and the reactive path.

sky_line added the

rendering

feature

labels

2026-10-11 05:15:22 +00:00

sky_line added this to the Slipstream project

2026-10-11 05:15:23 +00:00

sky_line commented

2026-10-11 05:15:23 +00:00

Author

Owner

Agent dispatched (Opus).

sky_line added the

status

in-progress

label

2026-10-11 05:15:23 +00:00

sky_line referenced this issue

2026-10-11 06:01:58 +00:00

DLSS upscaling + DLAA (native ML anti-aliasing) on NVIDIA RTX -- licence/Linux gate first #670

sky_line referenced this issue from a commit

2026-10-11 08:10:34 +00:00

#666: RT shadow denoiser everywhere -- projectors on the sun path, no TAA dependence, particles

sky_line referenced this issue from a commit

2026-10-11 08:10:34 +00:00

#664 round 3: default RenderVKRTShadowGrazingFade to 0 (A/B only)

sky_line commented

2026-10-11 08:53:33 +00:00

Author

Owner

Shipped in 1.3.0 Beta 3: projector shadows get the full RT shadow denoiser, the mask accumulates without TAA, and particles average their sun visibility.

Shipped in **1.3.0 Beta 3**: projector shadows get the full RT shadow denoiser, the mask accumulates without TAA, and particles average their sun visibility.

sky_line

2026-10-11 08:53:33 +00:00