DiagSplat · controlled RAW-noise experiment · 4 August 2026

NTF-v1 escapes the black trap
and wins the noisy RAW test

RawNeRF / sharpshadow with calibrated synthetic Bayer noise ×64. This report compares Vanilla 3DGS, the publicly reproducible Raw3DGS paper baseline, and the zero-gradient-safe NTF-v1 asinh transfer under an identical 31-train / 5-test protocol.

31 noisy RAW train views5 clean RAW test views30k iterations1008 × 756NTF-v1 · valid optimization

Clean target, Vanilla 3DGS, Raw3DGS paper baseline, and NTF-v1 rendering

NTF-v1 clean-test PSNR27.78 dB+1.67 dB over Vanilla

NTF-v1 clean-test SSIM0.868best of all three methods

NTF-v1 clean-test LPIPS0.180−29.9% versus Vanilla

NTF-v1 Gaussian count175,6899,783 fewer than Vanilla

先确认训练和测试数据

数据已拆成独立共享页面：完整列出全部 31 张 noisy training images、5 张 clean test images，以及 clean/noisy/noise-map protocol。以后相同数据的结果页都链接到同一个页面。

Open dataset page ↗

核心结论

NTF-v1 的提升是真实且一致的。相对 Vanilla，PSNR、SSIM 和 LPIPS 三项同时改善；逐视角 PSNR 在五张 clean held-out views 上全部高于 Vanilla。模型不再冻结为黑色，Gaussian 数量也没有异常坍缩。

这说明 NTF-v0 的失败主要来自实现边界条件：hard clamp 在 RawNeRF 黑色点云初始化处切断梯度。改为在零点具有有限非零梯度的 asinh toe 后，同一个核心 idea 可以正常优化。

Raw3DGS scope note
官方 Raw3DGS repository 仍未发布代码或 SID-pretrained noise extractor。本页的 “Raw3DGS paper baseline” 是论文明确描述、可复现的 HDR Scaffold-GS + RawNeRF scaling loss，不是完整 Raw3DGS 方法。它在该设置下过拟合到 34.17 dB train PSNR，而 clean test 只有约 26.4 dB。

最终定量结果

Method	PSNR ↑	SSIM ↑	LPIPS ↓	Stored representation	Interpretation
Vanilla 3DGS	26.1102	0.8387	0.2567	185,472 GS	valid noisy control
Raw3DGS paper baseline	26.3635	0.7526	0.3263	1,917,475 anchors	strong train/test overfit
NTF-v1 · asinh	27.7792	0.8676	0.1800	175,689 GS	best valid result

PSNR, SSIM, LPIPS, and representation-size comparison

Scaffold-GS stores anchors with ten learned offsets, so its 1.92M PLY vertices are anchors rather than directly equivalent standalone Gaussians. Even with this caveat, the growth is extreme and agrees with its large train/test gap.

逐视角 PSNR

Clean test view	Vanilla	Raw3DGS paper baseline	NTF-v1	NTF gain vs Vanilla
IMG_7120	24.78	27.98	25.98	+1.20
IMG_7128	25.77	23.09	27.01	+1.24
IMG_7136	25.51	26.32	27.68	+2.18
IMG_7144	26.65	26.33	28.66	+2.00
IMG_7152	27.84	28.09	29.57	+1.73

Five clean held-out views

每张测试图展示 clean target 与三种方法的统一 display-corrected rendering。Error map 是 camera-linear RGB mean absolute error，按 0.12 上限映射为 inferno 颜色。点击任何图像可放大。

可复现配置与 artifacts

Dataset	Shared RawNeRF sharpshadow data page · 31 noisy train / 5 clean test
Noise	Calibrated Poisson–Gaussian approximation in Bayer RAW · noise scale ×64
NTF-v1	asinh toe · effective shot 4.994896e−4 · read std 1.063407e−2 · target SNR 10 dB · low-slope ratio 0.125
Artifacts	Vanilla metrics · Raw3DGS paper baseline metrics · NTF-v1 metrics