Volumetric fog: crawling noise and speckling in fog light shafts (worse under DLSS); fog should read as fine mist / volume #1606

Open
opened 2026-11-18 10:53:26 +00:00 by sky_line · 2 comments
Owner

Volumetric fog (RenderVKVolumetricFog) shows crawling, temporally unstable noise. The clip shows it in a dark stairwell, where coloured light shafts come through a stained-glass window into the fog.

  • Under DLAA: there is still speckling in the shaft and fog area. That raw speckle is probably what DLSS picks up and amplifies.
  • In general: the fog has been noisy since it was introduced. It should read as a fine mist or a volume of cloud.
  • For comparison: particle fog (clean_pfog_*) already gives a smooth cloud with little noise.

Reported by the owner on 2026-11-18, RTX 5090, current main build (2c9514f131). Clip and frame attached.

Expected

  • Smooth fog with a stable image under a still and a moving camera, at native, DLAA and DLSS.
  • Light shafts with clean edges: no crawling, no speckle.
  • The fog should not rely on the upscaler to converge (see the "denoiser everywhere" rule: never rely on TAA/FSR/DLSS for convergence).

Where to look

  • indra/newview/vkgfx/llcleanvolfogvk.cpp
  • clean_light_froxel.comp, clean_medium.glsl, clean_medium_apply.comp
  • The fog's local-light shadows (RenderVKVolumetricFogLocalShadows)
  • Projector / stained-glass coloured shadows into fog
  • Sampling: per-froxel jitter, blue noise vs white noise, history reprojection and rejection, the reactive / transparency mask fed to DLSS and FSR, and froxel resolution vs shaft detail.

Done when

  • Before/after captures of this kind of scene show stable fog at native, DLAA and DLSS.
  • A measured noise metric is used (the RT bench noise probe, RenderVKNoiseStats, is a model).
  • There is no meaningful GPU cost increase.
Volumetric fog (`RenderVKVolumetricFog`) shows crawling, temporally unstable noise. The clip shows it in a dark stairwell, where coloured light shafts come through a stained-glass window into the fog. - **Under DLAA:** there is still speckling in the shaft and fog area. That raw speckle is probably what DLSS picks up and amplifies. - **In general:** the fog has been noisy since it was introduced. It should read as a fine mist or a volume of cloud. - **For comparison:** particle fog (`clean_pfog_*`) already gives a smooth cloud with little noise. Reported by the owner on 2026-11-18, RTX 5090, current main build (2c9514f131). Clip and frame attached. ## Expected - Smooth fog with a stable image under a still and a moving camera, at native, DLAA and DLSS. - Light shafts with clean edges: no crawling, no speckle. - The fog should not rely on the upscaler to converge (see the "denoiser everywhere" rule: never rely on TAA/FSR/DLSS for convergence). ## Where to look - `indra/newview/vkgfx/llcleanvolfogvk.cpp` - `clean_light_froxel.comp`, `clean_medium.glsl`, `clean_medium_apply.comp` - The fog's local-light shadows (`RenderVKVolumetricFogLocalShadows`) - Projector / stained-glass coloured shadows into fog - Sampling: per-froxel jitter, blue noise vs white noise, history reprojection and rejection, the reactive / transparency mask fed to DLSS and FSR, and froxel resolution vs shaft detail. ## Done when - Before/after captures of this kind of scene show stable fog at native, DLAA and DLSS. - A measured noise metric is used (the RT bench noise probe, `RenderVKNoiseStats`, is a model). - There is no meaningful GPU cost increase.
sky_line added this to the Slipstream project 2026-11-18 10:53:26 +00:00
Author
Owner

Root cause correction (owner, 2026-11-19): the crawling noise in the stairwell is RT reflections, not volumetric fog. With fog and god rays off, the noise remains.

  • It is worst on rough surfaces (plaster walls show vertical streaks and blotches) and least on smooth surfaces.
  • DLSS Ray Reconstruction was on (RenderVKDlssRayReconstruction=1).
  • New clip attached.

The fog-noise fix (fog-noise branch, merged) still measurably reduced the fog's own noise and stays. The reflection noise is handed to a Sol agent (brief: agents/sol_rt_reflection_noise_brief.md).

**Root cause correction (owner, 2026-11-19):** the crawling noise in the stairwell is **RT reflections, not volumetric fog**. With fog and god rays off, the noise remains. - It is worst on rough surfaces (plaster walls show vertical streaks and blotches) and least on smooth surfaces. - DLSS Ray Reconstruction was on (`RenderVKDlssRayReconstruction=1`). - New clip attached. The fog-noise fix (fog-noise branch, merged) still measurably reduced the fog's own noise and stays. The reflection noise is handed to a Sol agent (brief: agents/sol_rt_reflection_noise_brief.md).
Author
Owner

Second clip attached: the same stairwell at native, no DLSS (no RR). The rough-surface RT reflection noise is present without RR too, so it is in our own reflection path and denoiser, not RR amplification.

Second clip attached: the same stairwell at **native, no DLSS** (no RR). The rough-surface RT reflection noise is present without RR too, so it is in our own reflection path and denoiser, not RR amplification.
Sign in to join this conversation.
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
sky_line/Slipstream#1606
No description provided.