← Run Index
Run summary · P1 Phase F.1 A/B baseline
lego-phase-f1-baseline-5k
Phase F.1 baseline (use_simd_emit=false) re-take for A/B
partial pass
01
設定
- dataset
- /Users/otkrickey/dev/3dgs-workspace/datasets/nerf_synthetic/lego
- gt_convention
- Premultiplied
- iterations
- 5,000
- 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
31.63dB
Wallclock
1m 52s
Final splats
82,338
Final loss
1.448665e-2
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
Splat count — refine の発火と最終構成
05
中間ログ抜粋
| iter | 本 run loss |
|---|---|
| 1 | 5.9384e-1 |
| 500 | 8.6219e-2 |
| 1,000 | 5.5007e-2 |
| 1,500 | 3.2551e-2 |
| 2,000 | 1.9856e-2 |
| 2,500 | 1.7278e-2 |
| 3,000 | 1.6158e-2 |
| 3,500 | 1.5521e-2 |
| 4,000 | 1.5043e-2 |
| 4,500 | 1.4779e-2 |
| 5,000 | 1.4487e-2 |
06
関連 finding
この run を related_runs に含む finding (Positive / Negative どちらも)。
- negativePhase F.1 emit_pairs_simd + f16 forward gate flip — audit Tier 1 仮説 falsified、現規模で net regression / no improvementΔ PSNR ±0.13 dB (両者とも許容範囲、atomic/fp 順序由来)Δ +4.7% (emit_simd net regression) / +2.5% (f16 fwd noise 圏内)2026-05-25
- positivePhase G.3 SH-progressive — 5k smoke -14% は artifact、30k full は **quality improvement + 0.28 dB** に reframe、stacked + G.1 で **Pareto sweet spot** (Lego -61% wallclock + 0.15 dB)Δ PSNR +0.15 dB stacked (vs Phase D)、+0.28 dB 30k single (vs Phase D)、-0.12 dB 5k smoke (許容)Δ **-61% stacked** (vs Phase D)、-1.9% 30k single、-13.9% 5k smoke (artifact)2026-05-25