松田研究室
Run SummaryRun ID: chair-brushcompat-30k
Workspace: splat / commit 954c757
← Run Index
Run summary · P1.M5 multi-scene parity

chair-brushcompat-30k

P1.B+F Stage 2 multi-scene: chair brushcompat 30k

p1-b-fmulti-scenebrush-compatconvention-bridgepremultipliedstage-2-multichair
partial pass
01

設定

dataset
/Users/otkrickey/dev/3dgs-workspace/datasets/nerf_synthetic/chair
gt_convention
Premultiplied
iterations
30,000
seed
42
capacity
1,000,000 splats
sh_degree
3
loss
L1Ssim
lambda
0.200
ssim
window=7 sigma=1
backend.backward
Simd
backend.loss_path
Gpu
02

結果

Val PSNR
35.57dB
Wallclock
2h 3m 21s
Final splats
1,989,395
Final loss
3.537986e-3
03

3D ビューワー

学習結果 PLY を .splat 形式に変換し、antimatter15/splat の WebGL ビューワーで対話表示。

04

推移

Loss curve — 反復ごとの training loss
1e-2.61e-1.91e-1.31e-0.71e-0.16,00012k18k24k30kthis runloss
Splat count — refine の発火と最終構成
-94,002451,649997,3011,542,9532,088,6046,00012k18k24k30kthis runsplats
05

中間ログ抜粋

iter本 run loss
16.0968e-1
1,0004.5974e-2
2,0001.2954e-2
3,0008.3694e-3
4,0007.2443e-3
5,0006.6661e-3
6,0006.2644e-3
7,0005.9611e-3
8,0005.7036e-3
9,0005.4952e-3
10,0005.3203e-3
11,0005.1324e-3
12,0005.0089e-3
13,0004.8700e-3
14,0004.7824e-3
15,0004.6475e-3
16,0004.5026e-3
17,0004.3926e-3
18,0004.3076e-3
19,0004.2199e-3
20,0004.1367e-3
21,0004.0880e-3
22,0004.0071e-3
23,0003.9285e-3
24,0003.8700e-3
25,0003.8069e-3
26,0003.7607e-3
27,0003.7251e-3
28,0003.6354e-3
29,0003.6012e-3
30,0003.5380e-3
© 2026 大竹律輝 — 松田研究室