松田研究室
Run SummaryRun ID: ship-phase-g3-sh-progressive-30k
Workspace: splat / commit 4a0711c
← Run Index
Run summary · P1.M5 multi-scene parity (低 wallclock + 低 splats 版)

ship-phase-g3-sh-progressive-30k

P1.D multi-scene re-chain: ship brushcompat + opacity_decay 30k

p1-dmulti-scenebrush-compatpremultipliedopacity-decayphase-d-multiship
partial pass
01

設定

dataset
/Users/otkrickey/dev/3dgs-workspace/datasets/nerf_synthetic/ship
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
30.88dB
Wallclock
39m 46s
Final splats
495,160
Final loss
8.478498e-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-2.21e-1.71e-1.21e-0.61e-0.16,00012k18k24k30kthis runloss
Splat count — refine の発火と最終構成
-19,294115,462250,218384,974519,7306,00012k18k24k30kthis runsplats
05

中間ログ抜粋

iter本 run loss
15.8242e-1
1,0005.0317e-2
2,0002.5333e-2
3,0001.6769e-2
4,0001.3957e-2
5,0001.2804e-2
6,0001.2159e-2
7,0001.1851e-2
8,0001.1385e-2
9,0001.1230e-2
10,0001.0924e-2
11,0001.0905e-2
12,0001.0695e-2
13,0001.0693e-2
14,0001.0517e-2
15,0001.0484e-2
16,0009.5843e-3
17,0009.3151e-3
18,0009.1511e-3
19,0009.0796e-3
20,0008.9797e-3
21,0008.9027e-3
22,0008.8108e-3
23,0008.7697e-3
24,0008.7469e-3
25,0008.6254e-3
26,0008.6038e-3
27,0008.5663e-3
28,0008.5522e-3
29,0008.4677e-3
30,0008.4785e-3
© 2026 大竹律輝 — 松田研究室