GI: objects flash dark intermittently (irregular pace)

sky_line commented

2026-11-03 01:49:27 +00:00

Owner

Owner report: objects briefly flash dark, caused by GI. The flashes come at an irregular pace (not every frame, not a steady period), so it looks like an intermittent drop in the GI contribution on some objects rather than steady noise.

Seen on the current build (1.6 test at main fcbc3b846c, which carries the Beta 6 GI/stochastic work); check whether 1.5.0 Beta 6 shows it too, and whether it reproduces with GI off (then it isn't GI).

Candidates to rule in/out:

GI reuse / temporal history being rejected or reset for some pixels (disocclusion or validation failing intermittently), so the surface falls back to a dark first-bounce estimate for a frame.
Reservoir/cache slots evicted or aliased under load (irregular timing would match a budget or occupancy limit being hit now and then).
Stale GPU data during instance/BLAS updates (an object's GI lookup reading an invalid or zero entry for a frame while geometry or residency changes).
Interaction with DLSS Ray Reconstruction / upscaler history (owner runs DLSS Ultra Performance + RR).

Useful first data: a short capture with frame times plus the GI counters during the flashes, and whether the flashes line up with texture/mesh streaming activity or camera movement.

Owner report: objects briefly flash dark, caused by GI. The flashes come at an irregular pace (not every frame, not a steady period), so it looks like an intermittent drop in the GI contribution on some objects rather than steady noise. Seen on the current build (1.6 test at main fcbc3b846c, which carries the Beta 6 GI/stochastic work); check whether 1.5.0 Beta 6 shows it too, and whether it reproduces with GI off (then it isn't GI). Candidates to rule in/out: - GI reuse / temporal history being rejected or reset for some pixels (disocclusion or validation failing intermittently), so the surface falls back to a dark first-bounce estimate for a frame. - Reservoir/cache slots evicted or aliased under load (irregular timing would match a budget or occupancy limit being hit now and then). - Stale GPU data during instance/BLAS updates (an object's GI lookup reading an invalid or zero entry for a frame while geometry or residency changes). - Interaction with DLSS Ray Reconstruction / upscaler history (owner runs DLSS Ultra Performance + RR). Useful first data: a short capture with frame times plus the GI counters during the flashes, and whether the flashes line up with texture/mesh streaming activity or camera movement.

sky_line added the

bug

rendering

rt

labels

2026-11-03 01:49:27 +00:00

sky_line self-assigned this

2026-11-03 01:49:27 +00:00

sky_line added this to the Slipstream project

2026-11-03 01:49:27 +00:00

sky_line commented

2026-11-03 01:50:55 +00:00

Author

Owner

Owner, more observations (same session):

Without DLSS (no Ray Reconstruction), GI produces a massive amount of noise across the whole scene, so on non-RR paths the GI denoiser isn't converging. Per our rule the image must not rely on TAA/FSR/DLSS for convergence, so this is a bug in its own right.
With DLAA (native resolution, no upscaling), there are also shadow glitches.

So there may be three symptoms: dark flashes at an irregular pace (any mode?), heavy GI noise without RR, and shadow glitches under DLAA. Worth checking which of them share a cause, e.g. GI/shadow history or denoiser inputs that are only valid at the upscaled resolution or only when RR consumes them.

Owner, more observations (same session): - **Without DLSS** (no Ray Reconstruction), GI produces a massive amount of noise across the whole scene, so on non-RR paths the GI denoiser isn't converging. Per our rule the image must not rely on TAA/FSR/DLSS for convergence, so this is a bug in its own right. - **With DLAA** (native resolution, no upscaling), there are also shadow glitches. So there may be three symptoms: dark flashes at an irregular pace (any mode?), heavy GI noise without RR, and shadow glitches under DLAA. Worth checking which of them share a cause, e.g. GI/shadow history or denoiser inputs that are only valid at the upscaled resolution or only when RR consumes them.

sky_line referenced this issue from a commit

2026-11-03 02:26:11 +00:00

#1406 GI: keep the GI denoiser's history inside the emitter change spheres

sky_line referenced this issue from a commit

2026-11-03 02:37:22 +00:00

#1406 GI: bound each emitter change sphere by the emitter's own reach

sky_line referenced this issue

2026-11-04 08:29:32 +00:00

GI: alpha-blended windows flash brighter rapidly (GI off fixes it) #1438

sky_line referenced this issue from a commit

2026-11-04 11:30:43 +00:00

#1438 GI: glass no longer flashes brighter when a change sphere touches its cache cells