← Run Index
Run summary · docs/strategy/2026-08-next-phase-tasklist.md Tier2 A7 (refine の 3-pass 削減)
a7-12500-on-r3
A7 gate arm B (refine_scratch_persist = true, refine_direct_write = true, candidate): Lego 12.5k full run repeat 3
partial pass
01
設定
- dataset
- /Users/otkrickey/dev/3dgs-workspace/datasets/nerf_synthetic/lego
- gt_convention
- Premultiplied
- iterations
- 12,500
- 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
- backend.emit_pairs
- Original
- backend.forward
- F32
- backend.dispatch
- PerStage
- backend.radix_scan
- GpuHierarchical
- backend.radix_scatter
- SimdBallot
- backend.binning
- Resident
- backend.tile_sort
- GlobalRadix
- backend.submission
- Chained
- backend.target_cache
- true
- backend.shader_fast_math
- true
- os_build
- 26.6.2 (25G83)
- backend.ssim_clear
- HostMemset
- backend.raster_grad_clear
- HostMemset
- backend.backward_uniform
- HostBuffer
- backend.lazy_c3_scratch
- false
- backend.ssim_walk
- PerChannel
02
結果
Val PSNR
36.22dB
Wallclock
2m 4s
Final splats
407,210
Final loss
5.472030e-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
Splat count — refine の発火と最終構成
05
中間ログ抜粋
| iter | 本 run loss |
|---|---|
| 1 | 5.9385e-1 |
| 1,000 | 5.4771e-2 |
| 2,000 | 1.4153e-2 |
| 3,000 | 8.2763e-3 |
| 4,000 | 7.2409e-3 |
| 5,000 | 6.4613e-3 |
| 6,000 | 6.1193e-3 |
| 7,000 | 5.8548e-3 |
| 8,000 | 5.8013e-3 |
| 9,000 | 5.6394e-3 |
| 10,000 | 5.5991e-3 |
| 11,000 | 5.4734e-3 |
| 12,000 | 5.4381e-3 |
| 12,500 | 5.4720e-3 |