Systematic ladder test results. One entry per ladder step. Full plates and stats in each step's subfolder.
Diagnostic plate layout (4×3 grid):
| col 1 | col 2 | col 3 | col 4 | |
|---|---|---|---|---|
| r1 | render | accum raysTraced | GT-err Δ vs vanilla | noise Δ vs vanilla |
| r2 | frame level | accum maturity | accum mean | accum variance |
| r3 | accum coldmiss | frame qAHash | frame qBHash | frame probeSteps |
Both r1c3 and r1c9 use the same continuous bipolar ramp anchored at viridis(0) = dark purple for Δ = 0. Positive values (VisCache degraded / noisier) walk the full viridis palette (purple → blue → green → yellow); negative values (VisCache better / smoother) fade from purple toward black. Darker-than-purple = better; brighter-than-purple = worse. Plate labels report mean and per-pixel [min … max] signed %.
- r1c3 GT-err Δ = OkLabDistance(viscache, GT) − OkLabDistance(vanilla_xN, GT) at matched SPP — perceptual error vs ground truth, relative to same-SPP vanilla.
- r1c9 noise Δ = bilateral_noise(viscache LDR) − bilateral_noise(vanilla_xN LDR) at matched SPP — screen-space noise difference, independent of GT.
- Step 00 also emits per-SPP
{tag}_vanilla_gterr.png— vanilla_xN's absolute OkLab error vs GT — as the reference noise floor the viscache delta is measured against.
Vanilla PathTracer (no VisCache) at x1 / x16 / x4096 SPP. Error and ground-truth reference for downstream steps.
| x1 SPP | x16 SPP | x4096 SPP |
|---|---|---|
 |
 |
 |
PRESET_MINIMAL + RR_ADAPTIVE + FOOTPRINT_OFF, QUANT_MID, 1 logical frame, 1 spp. Sweep: 1×1 (artifact
baseline), 2×2 (+warmup 0/1/2), 4×4 (+warmup 0/1/8). Subframe gate disperses per-pixel cell writes across
N² Bayer sub-dispatches, breaking the tile-local first-writer-wins pattern while preserving warp coherence.
A subval control variant (pMin=2.0, bootThreshold/matureThreshold = 1 << 20) runs alongside — pins RR
to always-trace, isolating Bayer-dispatch + accumulation plumbing from cache-skip effects. noise_blob=0.00
across all N and scenes: the subframe path is bitwise-stable vs vanilla accumulation.
Insight — RR skip activity is cell-count-bound, not subframe-bound: on single-light scenes (1AreaLight,
1PointLight) 1×1 main sees ~2-4% RR skips at stock bootThreshold=8; multi-light scenes (3AreaLights,
32PointLights) stay at 100% trace because QUANT_MID's fine cells never reach maturity in 1 logical frame.
2×2 / 4×4 see no skips either — dispersion doesn't change the total per-cell sample count, only the order.
Insight — CV+RRR estimator variance shows up as noise in main-sweep plates: cells where 0.01 < p < 1
and we happen to trace (xi < p) feed mu + (V − mu)/p into the contribution. When p < 1, that's a
genuine unbiased estimator with variance, not a bug — shows up as modest noise_blob elevation in 1AL 2×2
(+54%) vs the subval plumbing baseline (0). Structural; not a plumbing issue.
Open points / possible improvements:
- Step 01's current cell size (
QUANT_MID) no longer exposes the cold-start tile artifact that was the original motivation for this step — cells are too fine to mature in 1 frame. Reintroducing a coarserQUANT_01variant (e.g. posA=0.5) would make the demo meaningful again on multi-light scenes. - Subval carries a small per-scene constant err_blob bias (+0.7–2.5%), invariant across N. Attributable to
Lr * visWeightvs bareLremission divergence and diagnostic UAV scheduling in theUSE_VISCACHE_VISIBILITYCHECK=1compiled shader. A true bitwise-vanilla control requires gating the whole NEE path on#ifdef, not a runtime flag.
Naive first pass: single level, uniform QUANT_SMALL, all 10 variants (pos_norm1 + pos_norm families), 4 scenes × x1 SPP. Warmup-write-only ON, footprint OFF, no cascade.
Best variant: pos_norm__dir1_dist1 — 39.6% rays traced (60.4% savings), 0.2% cold miss.
PRESET_MINIMAL + RR_ADAPTIVE. 4 quantization settings (qA→qD, posA 0.03→0.24 geometric) × 3 non-collapsed
norm-active B-side variants (pos, dir_dist1, dir_dist). 12 runs per scene.
Variant names: pos_norm__<bside>__<qtag>. Footprint still OFF.
Note (2026-04-14): The original step 03 included a fourth variant
dir_nearest(later renameddir_cleardist) that collapsed the distance bin and stored a per-cell distance signal for a short-ray visibility override. The concept was useable under the surface-target assumption but the complexity (second per-slot buffer, eviction-reset coupling, bin-homogeneity caveats, firefly risk on mispredictions) outweighed the expected gain over properdir_distbinning. Moved to future work — see paper conclusion, "Per-cell distance prior for collapsed-distance addressing". Step 03 reruns at 12 variants.
Fine B-side only × norm vs norm1, x1 and x16 SPP, 4 scenes each.
Finding: norm vs norm1 makes no measurable difference on convex geometry.
Decision: proceed with pos_norm only — uncollapsed normals handle thin-plate geometry
correctly; the CornellBox is too convex to expose the alias risk of norm1.
Best: pos_norm__pos at 23.4% mean rays x16 (76.6% savings).
Single level, footprintScale = 0 (pure bootThreshold gate, no footprint-aware trust). Sweeps bootThreshold over {4, 8, 16} with matureThreshold fixed. x1 / x4 SPP.
Key observation — depth gradient in maturity: the r1c4_accum_maturity column shows a clear depth gradient under all three threshold values — near-camera cells (large cellPixels, many parallel writes per dispatch) reach bootThreshold in one or two frames and read as "mature" (bright), while far cells (small cellPixels, sparse writes) stay immature (dark) much longer. This is the tile-boundary artifact that the footprint-aware trust gate is designed to fix: near cells look mature after one dispatch not because their estimates are actually stable, but because many threads wrote to the same cell within a single frame without producing spatial diversity.
Motivates the footprint addition (step 06): scaling the trust floor by log2(cellPixels) lifts near-camera cells' maturity bar so they have to collect writes across frames (where camera jitter provides genuine spatial diversity), not just within one dispatch. Far cells' trust threshold stays at bootThreshold as before.
Adds LEVELS_MULTI + auto-tuned cell sizes on top of step 05.
Cascade descent lets fine levels correct coarse-level early trust decisions.
Same as step 06 with THRESH_HIGH (bootThreshold=64, varThreshold=0.20 — higher than step 06's 8/0.10).
Isolates the effect of demanding more samples before trusting an entry.