GI noise near moving lights (point lights, probably projectors): worst at half GI resolution, colourful under DLSS #1087

New issue

Open

opened 2026-10-25 22:07:41 +00:00 by sky_line · 1 comment

sky_line commented

2026-10-25 22:07:41 +00:00

Owner

Owner, 2026-10-25 (dev fc780b19b3, RTX 5090, ray-traced GI on):

Noise on objects and avatars only near moving lights (scripted/animated point lights; probably projectors too). None under static lights.
Freezing the scene in that area (moving lights stop) removes the noise entirely; unfreezing brings it back.
Worst at half GI resolution (RenderVKGIScale < 1), with or without an upscaler.
At full GI resolution under DLSS (Ray Reconstruction): milder, colourful noise on objects.
Independent of the #1038 changes as far as tested.

Likely area: GI's temporal reuse does not react to light motion. When a light moves, cached / reused GI is stale but still accepted:

ReSTIR GI temporal reservoirs keep samples whose radiance came from the light's old position (no light-change validation; the "change sphere" logic of #985 covers geometry, does it cover lights?).
The radiance cache cells near the light keep old radiance (no invalidation / faster decay where a light moved).
The GI denoiser's history clamp / anti-lag doesn't see the change, then RR or the half-res upsample turns the mismatch into coloured noise.
At half resolution, fewer samples per pixel plus the upsample make each stale/fresh mismatch bigger.

Wanted: moving lights invalidate or fast-decay exactly the GI they affect (reservoirs whose path hit inside the light's moved volume, radiance-cache cells inside its old and new influence spheres, denoiser history via an anti-lag / light-change signal), leaving static areas fully temporally reused.

Owner, 2026-10-25 (dev fc780b19b3, RTX 5090, ray-traced GI on): - Noise on objects and avatars **only near moving lights** (scripted/animated point lights; probably projectors too). None under static lights. - **Freezing the scene** in that area (moving lights stop) removes the noise entirely; unfreezing brings it back. - Worst at **half GI resolution** (`RenderVKGIScale` < 1), with or without an upscaler. - At full GI resolution under DLSS (Ray Reconstruction): milder, **colourful** noise on objects. - Independent of the #1038 changes as far as tested. Likely area: GI's temporal reuse does not react to light motion. When a light moves, cached / reused GI is stale but still accepted: - ReSTIR GI temporal reservoirs keep samples whose radiance came from the light's old position (no light-change validation; the "change sphere" logic of #985 covers geometry, does it cover lights?). - The radiance cache cells near the light keep old radiance (no invalidation / faster decay where a light moved). - The GI denoiser's history clamp / anti-lag doesn't see the change, then RR or the half-res upsample turns the mismatch into coloured noise. - At half resolution, fewer samples per pixel plus the upsample make each stale/fresh mismatch bigger. Wanted: moving lights invalidate or fast-decay exactly the GI they affect (reservoirs whose path hit inside the light's moved volume, radiance-cache cells inside its old and new influence spheres, denoiser history via an anti-lag / light-change signal), leaving static areas fully temporally reused.

sky_line added the

rendering

bug

labels

2026-10-25 22:07:41 +00:00

sky_line added this to the Slipstream project

2026-10-25 22:07:41 +00:00

sky_line commented

2026-10-26 00:20:43 +00:00

Author

Owner

Owner test of dev 0979c17fa5: GI noise near the moving lights is much improved with the fix (9d30412d25). A small remainder seems to come from an animated fire (light + texture animation), which may be the same mechanism for animated emissive textures rather than moving light volumes -- to classify.

Owner test of dev 0979c17fa5: GI noise near the moving lights is much improved with the fix (9d30412d25). A small remainder seems to come from an animated fire (light + texture animation), which may be the same mechanism for animated emissive textures rather than moving light volumes -- to classify.

sky_line referenced this issue from a commit

2026-10-26 00:26:21 +00:00

#1087: GI under moving lights -- hold changed lights, clamp their reservoirs' M, fast-decay the cache

sky_line referenced this issue from a commit

2026-10-26 00:30:04 +00:00

#1087: GI under moving lights -- hold changed lights, clamp their reservoirs' M, fast-decay the cache

sky_line self-assigned this

2026-10-26 02:58:27 +00:00

sky_line referenced this issue from a commit

2026-10-26 20:49:28 +00:00

#1039 GI reuse occupancy: the reservoirs' changed-light tags in a stage of their own (CHG-TAG)

sky_line referenced this issue from a commit