Snapshots larger than the window are slightly blurry since #932 (no sharpening after the accumulation) #1093

New issue

Open

opened 2026-10-26 00:49:34 +00:00 by sky_line · 0 comments

sky_line commented

2026-10-26 00:49:34 +00:00

Owner

Users report snapshots a little blurrier in recent versions. Confirmed by a reporter's test: snapshots at the current window size are sharp; snapshots at other sizes (larger than the window) are soft.

Investigation (research/SNAPSHOT_BLUR.md, local): sizes above the window go through the jittered live accumulation (#873). The #932 grain fix (8abf303470) turned CAS sharpening off on the jittered samples (llcleanrenderervk.cpp ~30566, lens2[2] = 0 while capturing) and nothing sharpens the accumulated result instead; before #932 per-sample CAS partly offset the accumulator's softening. Contributing: the Gaussian reconstruction width (kAccSigma 0.5 output px), a fixed -0.5 texture mip bias while supersampling (2x wants about -1, 4x about -2), and bilinear-class gamma-space CPU scale-ups in the reduced-render fallbacks.

Fix without bringing back #932's grain: sharpen the accumulated mean once (linear light, before the output dither, strength from RenderCASSharpness), density-aware reconstruction width or a sharper kernel, mip bias by supersampling factor while a shot accumulates, Lanczos in linear light for the CPU scale-ups.

Users report snapshots a little blurrier in recent versions. Confirmed by a reporter's test: snapshots **at the current window size are sharp**; snapshots at **other sizes (larger than the window)** are soft. Investigation (research/SNAPSHOT_BLUR.md, local): sizes above the window go through the jittered live accumulation (#873). The #932 grain fix (8abf303470) turned CAS sharpening off on the jittered samples (`llcleanrenderervk.cpp` ~30566, `lens2[2] = 0` while capturing) and nothing sharpens the accumulated result instead; before #932 per-sample CAS partly offset the accumulator's softening. Contributing: the Gaussian reconstruction width (`kAccSigma` 0.5 output px), a fixed -0.5 texture mip bias while supersampling (2x wants about -1, 4x about -2), and bilinear-class gamma-space CPU scale-ups in the reduced-render fallbacks. Fix without bringing back #932's grain: sharpen the accumulated mean once (linear light, before the output dither, strength from RenderCASSharpness), density-aware reconstruction width or a sharper kernel, mip bias by supersampling factor while a shot accumulates, Lanczos in linear light for the CPU scale-ups.

sky_line added the

rendering

bug

labels

2026-10-26 00:49:34 +00:00

sky_line added this to the Slipstream project

2026-10-26 00:49:34 +00:00

sky_line referenced this issue from a commit

2026-10-26 02:36:20 +00:00

#1093: sharpen the accumulated snapshot once, density-adaptive splat width, mip bias while a shot accumulates

sky_line referenced this issue from a commit

2026-10-26 02:46:24 +00:00