DiagSplat · first low-light mechanism check · 2 August 2026

Noise-adaptive transfer functions for low-light 3DGS

Vanilla 3D Gaussian Splatting versus NTF-GS v0 on RealX3D Lowlight / Ujikintoki. Both models use the same 30 noisy training views, five clean held-out views, camera poses, initialization, and 30k-iteration schedule.

30 noisy train views 5 clean test views full resolution 30,000 iterations densification unchanged

Fewer Gaussians at 30k−16.7%862k → 718k

Rendered-image PSNR−0.21 dB21.242 → 21.030

LPIPS, lower is better−0.00350.4607 → 0.4572

Training time−7.9%17:55 → 16:30

Result in one sentence

NTF-GS v0 produces a materially smaller and slightly faster representation while preserving perceptual quality, but it does not yet beat vanilla 3DGS on PSNR.

This is a useful mechanism signal: the noise-adaptive objective suppresses some noise-driven Gaussian growth. It is not yet a complete low-light reconstruction win.

What the first run establishes The representation shrinks by 143,907 Gaussians without an SSIM or LPIPS regression. The next experiment should tune the transfer threshold and confidence strength, then measure local high-frequency recovery and Gaussian scale distributions.

Five clean held-out views

All images below come from the final 30k checkpoints. Click any image to inspect it at full resolution. Frame IDs correspond to the RealX3D test split.

Held-out view 1

0031.JPG

Clean ground truth for held-out frame 0031 — Clean GT

Vanilla 3DGS render for held-out frame 0031 — Vanilla 3DGS

NTF-GS v0 render for held-out frame 0031 — NTF-GS v0

Held-out view 2

0032.JPG

Clean ground truth for held-out frame 0032 — Clean GT

Vanilla 3DGS render for held-out frame 0032 — Vanilla 3DGS

NTF-GS v0 render for held-out frame 0032 — NTF-GS v0

Held-out view 3

0033.JPG

Clean ground truth for held-out frame 0033 — Clean GT

Vanilla 3DGS render for held-out frame 0033 — Vanilla 3DGS

NTF-GS v0 render for held-out frame 0033 — NTF-GS v0

Held-out view 4

0034.JPG

Clean ground truth for held-out frame 0034 — Clean GT

Vanilla 3DGS render for held-out frame 0034 — Vanilla 3DGS

NTF-GS v0 render for held-out frame 0034 — NTF-GS v0

Held-out view 5

0035.JPG

Clean ground truth for held-out frame 0035 — Clean GT

Vanilla 3DGS render for held-out frame 0035 — Vanilla 3DGS

NTF-GS v0 render for held-out frame 0035 — NTF-GS v0

Rendered-image metrics

Metrics are recomputed from the saved 30k renders with the standard Graphdeco evaluator. Bold values are better according to each metric.

Method	PSNR ↑	SSIM ↑	LPIPS ↓	Gaussians ↓
Vanilla 3DGS	21.2421	0.6289	0.4607	862,369
NTF-GS v0	21.0305	0.6295	0.4572	718,462
NTF − Vanilla	−0.2116	+0.0006	−0.0035	−143,907

The values printed during training differ slightly because the saved-image evaluator measures the final exported PNGs; this table is the consistent comparison used for the report.

Gaussian growth

NTF-GS remains smaller at both saved checkpoints. Densification code and thresholds are identical, so this difference is induced by the changed image-space gradients.

Vanilla 3DGS

862,369

NTF-GS v0

718,462

Checkpoint	Vanilla 3DGS	NTF-GS v0	Reduction
7k	667,333	576,578	−13.6%
30k	862,369	718,462	−16.7%

What the model actually sees

The physical low-light frames are extremely dark. For this controlled first run, each frame receives one scalar exposure gain estimated in linear RGB from its paired clean training reference. No spatial denoising or clean pixels are used as the optimization target.

Training frame 0001

gain 164.6×

Original physical low-light training image 0001 — Physical low-light input

Exposure-aligned noisy training image 0001 — Exposure-aligned noisy target

Clean paired reference image 0001 used only for scalar alignment — Clean reference · alignment only

Training frame 0015

gain 171.9×

Original physical low-light training image 0015 — Physical low-light input

Exposure-aligned noisy training image 0015 — Exposure-aligned noisy target

Clean paired reference image 0015 used only for scalar alignment — Clean reference · alignment only

Training frame 0030

gain 169.8×

Original physical low-light training image 0030 — Physical low-light input

Exposure-aligned noisy training image 0030 — Exposure-aligned noisy target

Clean paired reference image 0030 used only for scalar alignment — Clean reference · alignment only

Method difference

Vanilla 3DGS

Standard sRGB-domain L1 + SSIM objective
Standard gradient-driven densification
30k iterations at full image resolution

NTF-GS v0

Linear-radiance reconstruction
Two-slope logarithmic transfer function
Poisson–Gaussian SNR confidence weighting
Shot 0.01 · read 0.02 · threshold 0 dB · low-slope ratio 1/8
Densification itself remains unchanged

Interpretation and next check

Positive: the objective changes the densification trajectory in the desired direction and does not reduce SSIM or LPIPS on the clean test set.

Limitation: a 0.21 dB PSNR cost remains, and global image metrics cannot tell whether small noisy floaters or fine texture improved.

Next: sweep the SNR threshold and confidence gamma, render matched high-frequency crops, and report Gaussian scale/opacity histograms. A second-stage densification gate should only be added after selecting a loss-domain setting that closes the PSNR gap.

RealX3D Lowlight / Ujikintoki · 30 training views · 5 clean held-out views · experiment tag 20260802-first-v0