VSF3DGS / gate design
13 Nov 2026 · branch selflimit_gate

A self-limiting densification gate: R1 and R2

Low-light Gaussian splatting needs an allocation signal that sees dark detail, and a brake that stops it from fitting noise. We let an illumination-invariant score decide where to add Gaussians and a calibrated significance test decide whether the data support them. Growth then stops on its own once the residual is noise: no Gaussian cap and no per-dataset threshold tuning.

25–38 M
Gaussians grown by uncapped Raw3DGS on Dark3R (sRGB 12.7–13.6 dB)
30–100 %
of pure-noise Gaussians that pass the old gate
0.13 M
Gaussians where R1 stops by itself on chapel, at a higher PSNR than any capped baseline

1 · Native density control fails in both directions

Each host thresholds a screen-space gradient of its own photometric loss. In the dark that statistic is miscalibrated:

Our allocation score is relative L1 with a fixed scale λ = 1, so it sees dark structure like the same scene well exposed. It also amplifies dark noise, so a brake is required. The chart shows what happens with the old gate (R0), with a blind birth quota (Q) and with the new gates.

native (no cap) R0: old gate, λ = 1 R1 Q: 10k births / event R2
Number of Gaussians during training, Raw3DGS on Dark3R chapel, log scale. Native is the A800 uncapped run (sparse growth log; no growth after densification ends at 10k). R0 was stopped at step 6k after it passed 4 M. R1 and R2 flatten after about 3k steps; the step at 5k is the host's fixed injection of 100k extra points.
Show data

2 · Where vs whether

The gate leaves the host's clone/split/prune rules untouched. It only decides which candidates the host is allowed to densify. Both R1 and R2 use the same evidence. They differ only in how that evidence is combined with the host threshold.

Render μ noisy target y w = 1 / (μ + τ) Signal g (where) w · sign(μ − y) · ∂μ/∂u relative L1, λ = 1 Null probe h w · ε · ∂μ/∂u ε = coloured ±1 noise Window sums S_g = Σ‖g‖², S_h = Σ‖h‖² n_eff = S_h² / Σ‖h‖⁴ R1 · two tests host score ḡ ≥ τ_host AND z = S_g / S_h ≥ q R2 · one debiased score s = √( max(0, S_g − q S_h) / n ) s ≥ τ_host replaces the host statistic q = F⁻¹(1 − α/N; n_eff, n_eff) every 100 steps → host clone / split
One densification window (100 training views). The probe costs one extra backward pass through the same render graph; no CUDA change.

3 · The test

For Gaussian $i$ and training view $t$, let $u_i$ be its projected 2-D centre and $\mu_p$ the clamped rendered value at pixel $p$. The allocation weight is relative: $w_p = 1/(\mu_p + \tau)$, $\tau = 10^{-3}$.

signal: the host-routed allocation gradient (λ = 1)
$$ g_i^t \;=\; \frac{\partial}{\partial u_i}\,\frac{1}{|P|}\sum_p \operatorname{sg}(w_p)\,\lvert \mu_p - y_p\rvert \;=\; \frac{1}{|P|}\sum_p w_p\,\operatorname{sign}(\mu_p - y_p)\,\frac{\partial \mu_p}{\partial u_i} $$
null probe: the same sum with the residual signs replaced by noise of the data's spatial correlation
$$ h_i^t \;=\; \frac{1}{|P|}\sum_p w_p\,\varepsilon_p\,\frac{\partial \mu_p}{\partial u_i}, \qquad \varepsilon = \operatorname{sign}(k * \xi),\;\; \xi \sim \mathcal N(0, I) $$
evidence over the views $t$ of one densification window
$$ S_g = \sum_t \lVert g_i^t\rVert^2,\qquad S_h = \sum_t \lVert h_i^t\rVert^2,\qquad z_i = \frac{S_g}{S_h},\qquad n_{\text{eff}} = \frac{S_h^2}{\sum_t \lVert h_i^t\rVert^4} $$
critical value: F test, Bonferroni over the $N$ Gaussians tested at this event (α = 1 expected false birth)
$$ q_i \;=\; F^{-1}_{n_{\text{eff}},\,n_{\text{eff}}}\!\left(1 - \frac{\alpha}{N}\right) $$

If the residual signs inside a footprint are noise, $g$ and $h$ have the same distribution and $z \approx 1$. A real misfit, such as an edge in the wrong place or a missing texture, makes the signs coherent with the footprint's sensitivity, and the signal energy adds up. Panel C shows the subtle case: demosaiced noise is spatially correlated, so it looks "coherent" to a pixel-independent null.

Residual sign maps under a Gaussian footprint for structure, independent noise and correlated noise, with the resulting z values
render below target render above target Solid and dashed rings: 1σ and 2σ of the footprint. The z values are energy ratios for this footprint, averaged over many noise draws. The sensitivity is odd, so a left/right sign split (A) adds up. Uncorrelated noise (B) gives z ≈ 1. Correlated noise (C) gives z ≈ 4 against an i.i.d. null, which is a false birth; against a null with the same correlation it gives z ≈ 1.

4 · Calibrating the null to the camera pipeline

The old gate (R0) used pixel-independent ±1 signs. Real low-light images are demosaiced and resampled, so neighbouring noise samples are correlated. We measure the sign autocorrelation once per dataset from a pure-noise residual: Dark3R uses low − 0.25 × long-exposure pairs, RawNeRF ×64 uses the strong64 − native difference. From it we build a small kernel $k$ such that $\operatorname{sign}(k*\xi)$ has the same autocorrelation.

Real Dark3R noise signs
Real Dark3R noise · lag-1 autocorr. 0.27
Independent signs
Old null (i.i.d. ±1) · 0.02
Coloured signs
New null (coloured ±1) · 0.31

Same 40 × 40 crop size, one sign per pixel. The real patch is a flat region of a chapel training frame.

R0 rule: i.i.d. null, z ≥ 1 + 3√(2/n) R1 test: coloured null + F, α = 10⁻² per test
Share of Gaussians that pass the gate when the residual is pure noise. The test uses held-out frames not used to fit $k$, a footprint of σ = 2 px and n views per window. In training, α/N is about 10⁻⁵, so R1's false-birth rate is lower still.
Show data

5 · R1 and R2

R1 · score and test recommended default

$$ \text{densify } i \iff \bar g_i \ge \tau_{\text{host}} \;\wedge\; z_i \ge q_i $$

$\bar g_i = \frac1n\sum_t \lVert g_i^t\rVert$ is the host's own statistic. The allocation score at λ = 1 says where; the test says whether.

  • Keeps the decoupling explicit: allocation and detection stay separate, which is the paper's story.
  • α has a meaning: expected false births per event. Its false-positive rate is verified (§4).
  • A filter on the host rule. The host statistic and threshold are unchanged.

R2 · one debiased score

$$ s_i = \sqrt{\max\!\big(0,\; S_g - q_i S_h\big)/n_i}\;\ge\;\tau_{\text{host}} $$

The noise floor $q_i S_h$ is subtracted before thresholding. Only the excess, structured energy is compared with the host threshold.

  • Elegant: a single number per Gaussian, and the gate disappears into the score.
  • But it replaces the host statistic (mean norm → RMS excess), and α and τ_host interact.
  • Keep as an ablation: "the same evidence, combined differently".
Decision rule, fixed in advance: use R1 unless R2 is better by ≥ 0.3 dB mean sRGB PSNR on both Dark3R and RawNeRF, with no LPIPS loss, over seeds. So far they are tied on Dark3R. R2 is ahead on RawNeRF, where there is a single test view and the margin is within the noise.

6 · Why it stops by itself

Early in training the model is under-fit: residual signs are coherent, z is large and the gate opens. As Gaussians absorb the structure, the residual becomes noise, z falls towards 1, below the critical value, and births stop. False births are bounded by α per event, about 95 over a whole run. The spike at 5k is the host's injection of 100k random points: they are briefly under-fit, the gate reopens, then closes again.

median z of candidates critical value q (median)
R1, Raw3DGS chapel: evidence vs threshold at each densify event (log scale).
R0 births R1 births
Gaussians allowed to clone/split per event (log scale). R0 keeps admitting tens of thousands.
Show data

7 · First numbers

Screening run: Raw3DGS, 15k iterations, seed 42. Each cell gives sRGB PSNR / sRGB LPIPS / final #Gaussians. The reference columns use a 100k Gaussian cap, which the historical protocol needed. R1 and R2 run with no cap at all.

scenenative @100k capold SN_G @100k capR1 · no capR2 · no capQ · 10k / event

Uncapped native at 30k on the A800: chapel 12.67, statue 13.55, office64 10.32 dB. Full 30k runs with R1, R2, birth-matched controls and α sweeps are in progress: Raw3DGS and LE3D on the A800, 3DGS and Lumi-GS on iridium. The previous seed-42 report for 3DGS and Lumi-GS is here.

8 · Open points