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

hotdog-phase-g3-sh-progressive-30k

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

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

設定

dataset
/Users/otkrickey/dev/3dgs-workspace/datasets/nerf_synthetic/hotdog
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
37.37dB
Wallclock
28m 17s
Final splats
310,045
Final loss
2.343549e-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-2.71e-2.11e-1.51e-0.91e-0.36,00012k18k24k30kthis runloss
Splat count — refine の発火と最終構成
-10,14974,311158,772243,232327,6926,00012k18k24k30kthis runsplats
05

中間ログ抜粋

iter本 run loss
14.3290e-1
1,0001.8886e-2
2,0006.9609e-3
3,0004.3936e-3
4,0003.9338e-3
5,0003.5234e-3
6,0003.2906e-3
7,0003.1358e-3
8,0003.0796e-3
9,0002.9833e-3
10,0002.9097e-3
11,0002.8179e-3
12,0002.8351e-3
13,0002.8119e-3
14,0002.8073e-3
15,0002.7594e-3
16,0002.6536e-3
17,0002.6074e-3
18,0002.5329e-3
19,0002.5036e-3
20,0002.5280e-3
21,0002.5151e-3
22,0002.5014e-3
23,0002.4548e-3
24,0002.4207e-3
25,0002.4385e-3
26,0002.4328e-3
27,0002.3726e-3
28,0002.3487e-3
29,0002.3661e-3
30,0002.3435e-3
© 2026 大竹律輝 — 松田研究室