松田研究室
Run Index285 runs · 2026-04-24 〜 2026-09-23
workspaces: splat
← Document Index
Experiment index

Apple Silicon 3DGS
実験 run 一覧

各 run の詳細サマリ HTML へのハブ。verdict / PSNR / wallclock / splat 数で一目で比較。

Total runs
285
Pass
0
Partial
281
Rejected
4
01

Run 一覧

verdict
tag
search
RunVerdict PSNR Wallclock Splats Date ▼
brush-sameday-lego-30kbrush 同日比較用 Lego 30k (2026-08-31-final-lego-30k.toml の run 名・run_dir・日付のみ変更した複製、backend default)
p3-final-stackmsplat-paritylego-30kfinal-stackheadline
partial36.42 dB6m 32s485,8252026-09-23
c1-drums-12500-off-profdrums (998k splats) per-kernel gpu_ms share 計測。Lego v2 (p3-kernel-share-post-s3 §6) の測定 regime を drums dataset で再現し、splat 数 regime の違いが share 構造に影響するかを判定する
p3-kernel-share-drumsmulti-scenedrumsper-kernel-gpu-msgpu-gate
partial24.52 dB4m 9s748,7082026-08-31
final-lego-30k最終スタック Lego 30k headline (msplat 30K 同一 iter 数比較、backend default)
p3-final-stackmsplat-paritylego-30kfinal-stackheadline
partial36.46 dB6m 53s490,8662026-08-31
final-lego-30k-r2最終スタック Lego 30k headline n=2 目 (msplat 30K 同一 iter 数比較、backend default)
p3-final-stackmsplat-paritylego-30kfinal-stackheadline
partial36.44 dB8m 2s488,4912026-08-31
finalv3-12500-r1thermal/DVFS drift 検証: 10 分 cooldown 後の Lego 12.5k full run repeat 1/2 (finalv2 と同一 config)
p3-final-stackm7gpu-gatelegofinal-stackthermal-drift
partial36.33 dB2m 4s405,7162026-08-31
finalv3-12500-r2thermal/DVFS drift 検証: 10 分 cooldown 後の Lego 12.5k full run repeat 2/2 (finalv2 と同一 config)
p3-final-stackm7gpu-gatelegofinal-stackthermal-drift
partial36.18 dB2m 8s409,5832026-08-31
s4b-12500-on-rerundrift 切り分け: 2026-08-30-s4b-12500-on-r1.toml と完全同一 config を run 名のみ変えて再実行 (§8 フォローアップ)
p3-final-stackdrift-checkgpu-gatelegocontrol-rerun
partial36.22 dB2m 5s409,6062026-08-31
a7-12500-off-r1A7 gate arm A (refine_scratch_persist = false, refine_direct_write = false, baseline): Lego 12.5k full run repeat 1
p3-refine-3passa7gpu-gatelego
partial36.29 dB2m 8s407,9692026-08-30
a7-12500-off-r2A7 gate arm A (refine_scratch_persist = false, refine_direct_write = false, baseline): Lego 12.5k full run repeat 2
p3-refine-3passa7gpu-gatelego
partial36.23 dB2m 6s405,5302026-08-30
a7-12500-off-r3A7 gate arm A (refine_scratch_persist = false, refine_direct_write = false, baseline): Lego 12.5k full run repeat 3
p3-refine-3passa7gpu-gatelego
partial36.15 dB2m 7s407,8002026-08-30
a7-12500-on-r1A7 gate arm B (refine_scratch_persist = true, refine_direct_write = true, candidate): Lego 12.5k full run repeat 1
p3-refine-3passa7gpu-gatelego
partial36.26 dB2m 5s410,5032026-08-30
a7-12500-on-r2A7 gate arm B (refine_scratch_persist = true, refine_direct_write = true, candidate): Lego 12.5k full run repeat 2
p3-refine-3passa7gpu-gatelego
partial36.25 dB2m 4s408,4352026-08-30
a7-12500-on-r3A7 gate arm B (refine_scratch_persist = true, refine_direct_write = true, candidate): Lego 12.5k full run repeat 3
p3-refine-3passa7gpu-gatelego
partial36.22 dB2m 4s407,2102026-08-30
p3-backward-audit-checkpoint-12500M8 atomic spread の splat bench --parity / replay bench 用 checkpoint (12500 iter, 実 scale)
p3-backward-auditcheckpoint-gen
partial36.20 dB2m 4s408,4692026-08-30
c1-12500-off-profM1 C1: per-kernel gpu_ms 再計測 (submission=per_kernel / forward_submission=per_stage / pipelining=off、post-S2/S3 kernel)
p3-kernel-share-post-s3m1gpu-gatelego
partial36.09 dB3m 25s408,0062026-08-30
c1v2-12500-off-profM7: per-kernel gpu_ms 再々計測 v2 (submission=per_kernel / forward_submission=per_stage / pipelining=off、M4 radix simd_ballot + A7 refine 3-pass 後の kernel)
p3-kernel-share-post-s3m7gpu-gatelegov2
partial36.20 dB2m 54s404,6522026-08-30
c1v3-12500-off-profM7: per-kernel gpu_ms 再々計測 v2 (submission=per_kernel / forward_submission=per_stage / pipelining=off、M4 radix simd_ballot + A7 refine 3-pass 後の kernel)
p3-kernel-share-post-s3m7gpu-gatelegov2
partial36.19 dB2m 54s409,7022026-08-30
c2-12500-depth2-profC2 計装確認: Lego 12.5k full run (profile=true, pipelining=depth2)。gpu_gap_report で残差 4.54s を cb_gap/drvproc/handoff/host_lead に分解する
c2gap-reportgpu-gatelego
partial36.07 dB2m 4s406,6512026-08-30
final-12500-r1M7 最終スタック独立再計測 (全採用施策 default ON): Lego 12.5k full run repeat 1/3
p3-final-stackm7gpu-gatelegofinal-stack
partial36.13 dB2m 11s407,7542026-08-30
final-12500-r2M7 最終スタック独立再計測 (全採用施策 default ON): Lego 12.5k full run repeat 2/3
p3-final-stackm7gpu-gatelegofinal-stack
partial36.20 dB2m 3s405,3332026-08-30
final-12500-r3M7 最終スタック独立再計測 (全採用施策 default ON): Lego 12.5k full run repeat 3/3
p3-final-stackm7gpu-gatelegofinal-stack
partial36.23 dB2m 4s408,7562026-08-30
final-12500-r4M7 最終スタック独立再計測 (全採用施策 default ON): Lego 12.5k full run repeat 4 (discriminator, r1 の sampler window 欠測を切り分けるための追加 run)
p3-final-stackm7gpu-gatelegofinal-stack
partial36.22 dB2m 5s406,9942026-08-30
finalv2-12500-r1M7 最終スタック独立再計測 (全採用施策 default ON): Lego 12.5k full run repeat 1/3 (v2, post-S4b flip)
p3-final-stackm7gpu-gatelegofinal-stack
partial36.28 dB2m 12s406,7732026-08-30
finalv2-12500-r2M7 最終スタック独立再計測 (全採用施策 default ON): Lego 12.5k full run repeat 2/3 (v2, post-S4b flip)
p3-final-stackm7gpu-gatelegofinal-stack
partial36.27 dB2m 11s406,9902026-08-30
finalv2-12500-r3M7 最終スタック独立再計測 (全採用施策 default ON): Lego 12.5k full run repeat 3/3 (v2, post-S4b flip)
p3-final-stackm7gpu-gatelegofinal-stack
partial36.21 dB2m 8s405,9462026-08-30
m4-12500-off-r1M4/A5 gate arm A (radix_scatter = linear_rank): Lego 12.5k full run repeat 1
p3-radix-ballotm4a5gpu-gatelego
partial36.20 dB2m 30s408,9282026-08-30
m4-12500-off-r2M4/A5 gate arm A (radix_scatter = linear_rank): Lego 12.5k full run repeat 2
p3-radix-ballotm4a5gpu-gatelego
partial36.02 dB2m 29s411,0492026-08-30
m4-12500-off-r3M4/A5 gate arm A (radix_scatter = linear_rank): Lego 12.5k full run repeat 3
p3-radix-ballotm4a5gpu-gatelego
partial36.15 dB2m 29s408,1132026-08-30
m4-12500-on-r1M4/A5 gate arm B (radix_scatter = simd_ballot): Lego 12.5k full run repeat 1
p3-radix-ballotm4a5gpu-gatelego
partial36.15 dB2m 8s408,4082026-08-30
m4-12500-on-r2M4/A5 gate arm B (radix_scatter = simd_ballot): Lego 12.5k full run repeat 2
p3-radix-ballotm4a5gpu-gatelego
partial36.06 dB2m 6s409,0742026-08-30
m4-12500-on-r3M4/A5 gate arm B (radix_scatter = simd_ballot): Lego 12.5k full run repeat 3
p3-radix-ballotm4a5gpu-gatelego
partial36.22 dB2m 7s406,7702026-08-30
m5a-12500-off-r1M5a SSIM float4 gate arm A (baseline, per_channel): Lego 12.5k full run repeat 1
p3-ssim-float4m5agpu-gatelego
partial36.26 dB2m 8s407,3992026-08-30
m5a-12500-off-r2M5a SSIM float4 gate arm A (baseline, per_channel): Lego 12.5k full run repeat 2
p3-ssim-float4m5agpu-gatelego
partial36.20 dB2m 7s408,1722026-08-30
m5a-12500-off-r3M5a SSIM float4 gate arm A (baseline, per_channel): Lego 12.5k full run repeat 3
p3-ssim-float4m5agpu-gatelego
partial36.18 dB2m 5s405,1212026-08-30
m5a-12500-on-r1M5a SSIM float4 gate arm B (float4_rcp 候補): Lego 12.5k full run repeat 1
p3-ssim-float4m5agpu-gatelego
partial36.16 dB2m 13s404,9012026-08-30
m5a-12500-on-r2M5a SSIM float4 gate arm B (float4_rcp 候補): Lego 12.5k full run repeat 2
p3-ssim-float4m5agpu-gatelego
partial36.17 dB2m 4s407,6572026-08-30
m5a-12500-on-r3M5a SSIM float4 gate arm B (float4_rcp 候補): Lego 12.5k full run repeat 3
p3-ssim-float4m5agpu-gatelego
partial36.07 dB2m 3s408,8262026-08-30
quiet-12500-r1p3-final-stack-v2 §13 静穏条件再計測: control config (s4b-12500-on-r1) の byte-identical 複製、repeat 1/3
p3-s4bs4bgpu-gatelegossim-reduce
partial36.17 dB2m 4s406,8502026-08-30
quiet-12500-r2p3-final-stack-v2 §13 静穏条件再計測: control config (s4b-12500-on-r2) の byte-identical 複製、repeat 2/3
p3-s4bs4bgpu-gatelegossim-reduce
partial36.21 dB2m 7s407,9462026-08-30
quiet-12500-r3p3-final-stack-v2 §13 静穏条件再計測: control config (s4b-12500-on-r3) の byte-identical 複製、repeat 3/3
p3-s4bs4bgpu-gatelegossim-reduce
partial36.19 dB2m 6s407,5532026-08-30
re200-12500-off-r1refine.every=100 (baseline) A/B arm A: Lego 12.5k full run repeat 1/3
p3-refine-every-200re200gpu-gatelegohyperparameter
partial36.13 dB2m 5s407,6642026-08-30
re200-12500-off-r2refine.every=100 (baseline) A/B arm A: Lego 12.5k full run repeat 2/3
p3-refine-every-200re200gpu-gatelegohyperparameter
partial36.14 dB2m 12s406,5542026-08-30
re200-12500-off-r3refine.every=100 (baseline) A/B arm A: Lego 12.5k full run repeat 3/3
p3-refine-every-200re200gpu-gatelegohyperparameter
partial36.10 dB2m 4s407,9202026-08-30
re200-12500-on-r1refine.every=200 (candidate) A/B arm B: Lego 12.5k full run repeat 1/3
p3-refine-every-200re200gpu-gatelegohyperparameter
partial35.77 dB1m 33s266,4742026-08-30
re200-12500-on-r2refine.every=200 (candidate) A/B arm B: Lego 12.5k full run repeat 2/3
p3-refine-every-200re200gpu-gatelegohyperparameter
partial35.74 dB1m 33s267,9592026-08-30
re200-12500-on-r3refine.every=200 (candidate) A/B arm B: Lego 12.5k full run repeat 3/3
p3-refine-every-200re200gpu-gatelegohyperparameter
partial35.70 dB1m 34s265,5402026-08-30
rnd-12500-off-r1C2 policy 4 (refine_gpu_snapshot) gate arm A (baseline, refine_gpu_snapshot = false): Lego 12.5k full run repeat 1/3
p3-refine-nondrainc2refine-gpu-snapshotgpu-gatelego
partial36.11 dB2m 3s408,7982026-08-30
rnd-12500-off-r2C2 policy 4 (refine_gpu_snapshot) gate arm A (baseline, refine_gpu_snapshot = false): Lego 12.5k full run repeat 2/3
p3-refine-nondrainc2refine-gpu-snapshotgpu-gatelego
partial36.19 dB2m 2s405,9762026-08-30
rnd-12500-off-r3C2 policy 4 (refine_gpu_snapshot) gate arm A (baseline, refine_gpu_snapshot = false): Lego 12.5k full run repeat 3/3
p3-refine-nondrainc2refine-gpu-snapshotgpu-gatelego
partial36.11 dB2m 2s407,5622026-08-30
rnd-12500-on-r1C2 policy 4 (refine_gpu_snapshot) gate arm B (candidate, refine_gpu_snapshot = true): Lego 12.5k full run repeat 1/3
p3-refine-nondrainc2refine-gpu-snapshotgpu-gatelego
partial36.18 dB2m 2s405,2422026-08-30
rnd-12500-on-r2C2 policy 4 (refine_gpu_snapshot) gate arm B (candidate, refine_gpu_snapshot = true): Lego 12.5k full run repeat 2/3
p3-refine-nondrainc2refine-gpu-snapshotgpu-gatelego
partial36.20 dB2m 2s406,6882026-08-30
rnd-12500-on-r3C2 policy 4 (refine_gpu_snapshot) gate arm B (candidate, refine_gpu_snapshot = true): Lego 12.5k full run repeat 3/3
p3-refine-nondrainc2refine-gpu-snapshotgpu-gatelego
partial36.20 dB2m 2s409,9492026-08-30
rnd-12500-prof-offC2 policy 4 (refine_gpu_snapshot) gate arm A (baseline, refine_gpu_snapshot = false): Lego 12.5k full run repeat 1/3
p3-refine-nondrainc2refine-gpu-snapshotgpu-gatelego
partial36.24 dB2m 3s409,6112026-08-30
rnd-12500-prof-onC2 policy 4 (refine_gpu_snapshot) gate arm B (candidate, refine_gpu_snapshot = true): Lego 12.5k full run repeat 1/3
p3-refine-nondrainc2refine-gpu-snapshotgpu-gatelego
partial36.10 dB2m 9s407,3192026-08-30
s4-12500-off-r1C2/S4 gate off arm (host_schedule=forward_shadow, 現行 default): Lego 12.5k full run repeat 1/3
p3-final-stackm7gpu-gatelegofinal-stack
partial36.16 dB2m 9s408,4152026-08-30
s4-12500-off-r2C2/S4 gate off arm (host_schedule=forward_shadow, 現行 default): Lego 12.5k full run repeat 2/3
p3-final-stackm7gpu-gatelegofinal-stack
partial36.15 dB2m 5s408,7492026-08-30
s4-12500-off-r3C2/S4 gate off arm (host_schedule=forward_shadow, 現行 default): Lego 12.5k full run repeat 3/3
p3-final-stackm7gpu-gatelegofinal-stack
partial36.24 dB2m 4s404,0032026-08-30
s4-12500-off-r4C2/S4 gate off arm (host_schedule=forward_shadow, 現行 default): Lego 12.5k full run repeat 4/3 (r1 retry, load contamination)
p3-final-stackm7gpu-gatelegofinal-stack
partial36.13 dB2m 5s408,6582026-08-30
s4-12500-on-r1C2/S4 gate on arm (host_schedule=backward_shadow): Lego 12.5k full run repeat 1/3
p3-final-stackm7gpu-gatelegofinal-stack
partial36.18 dB2m 8s403,2222026-08-30
s4-12500-on-r2C2/S4 gate on arm (host_schedule=backward_shadow): Lego 12.5k full run repeat 2/3
p3-final-stackm7gpu-gatelegofinal-stack
partial36.16 dB2m 6s408,0112026-08-30
s4-12500-on-r3C2/S4 gate on arm (host_schedule=backward_shadow): Lego 12.5k full run repeat 3/3
p3-final-stackm7gpu-gatelegofinal-stack
partial36.21 dB2m 8s409,3182026-08-30
s4-12500-on-r4C2/S4 gate on arm (host_schedule=backward_shadow): Lego 12.5k full run repeat 4/3 (r1 retry, load contamination)
p3-final-stackm7gpu-gatelegofinal-stack
partial36.22 dB2m 21s407,5802026-08-30
s4b-12500-off-r1S4b gate off arm (ssim_reduce=host, 現行 default): Lego 12.5k full run repeat 1/3
p3-s4bs4bgpu-gatelegossim-reduce
partial36.12 dB2m 5s410,3562026-08-30
s4b-12500-off-r2S4b gate off arm (ssim_reduce=host, 現行 default): Lego 12.5k full run repeat 2/3
p3-s4bs4bgpu-gatelegossim-reduce
partial36.15 dB2m 4s407,0842026-08-30
s4b-12500-off-r3S4b gate off arm (ssim_reduce=host, 現行 default): Lego 12.5k full run repeat 3/3
p3-s4bs4bgpu-gatelegossim-reduce
partial36.10 dB2m 4s409,3002026-08-30
s4b-12500-on-r1S4b gate on arm (ssim_reduce=gpu): Lego 12.5k full run repeat 1/3
p3-s4bs4bgpu-gatelegossim-reduce
partial36.21 dB2m 4s407,3782026-08-30
s4b-12500-on-r2S4b gate on arm (ssim_reduce=gpu): Lego 12.5k full run repeat 2/3
p3-s4bs4bgpu-gatelegossim-reduce
partial36.18 dB2m 2s405,5682026-08-30
s4b-12500-on-r3S4b gate on arm (ssim_reduce=gpu): Lego 12.5k full run repeat 3/3
p3-s4bs4bgpu-gatelegossim-reduce
partial36.14 dB2m 2s406,0922026-08-30
t1-12500-off-r1M2 Tier1 gate arm A (全 flag OFF、従来動作): Lego 12.5k full run repeat 1
p3-tier1-bit-exactm2gpu-gatelego
partial36.20 dB2m 30s409,1212026-08-30
t1-12500-off-r2M2 Tier1 gate arm A (全 flag OFF、従来動作): Lego 12.5k full run repeat 2
p3-tier1-bit-exactm2gpu-gatelego
partial36.19 dB2m 37s407,5052026-08-30
t1-12500-off-r3M2 Tier1 gate arm A (全 flag OFF、従来動作): Lego 12.5k full run repeat 3
p3-tier1-bit-exactm2gpu-gatelego
partial36.22 dB2m 37s409,5372026-08-30
t1-12500-on-r1M2 Tier1 gate arm B (全 flag ON、A1-A4 候補): Lego 12.5k full run repeat 1
p3-tier1-bit-exactm2gpu-gatelego
partial36.16 dB2m 28s408,5162026-08-30
t1-12500-on-r2M2 Tier1 gate arm B (全 flag ON、A1-A4 候補): Lego 12.5k full run repeat 2
p3-tier1-bit-exactm2gpu-gatelego
partial36.10 dB2m 26s405,1752026-08-30
t1-12500-on-r3M2 Tier1 gate arm B (全 flag ON、A1-A4 候補): Lego 12.5k full run repeat 3
p3-tier1-bit-exactm2gpu-gatelego
partial36.16 dB2m 28s409,6332026-08-30
t1v2-12500-off-r1M2 Tier1 再 gate v2 arm A (全 flag OFF、従来動作): Lego 12.5k full run repeat 1 (n=5, radix ballot default 上)
p3-tier1-bit-exactm2gpu-gatelegot1v2n5-regate
partial36.20 dB2m 7s408,9362026-08-30
t1v2-12500-off-r2M2 Tier1 再 gate v2 arm A (全 flag OFF、従来動作): Lego 12.5k full run repeat 2 (n=5, radix ballot default 上)
p3-tier1-bit-exactm2gpu-gatelegot1v2n5-regate
partial36.22 dB2m 6s407,0002026-08-30
t1v2-12500-off-r3M2 Tier1 再 gate v2 arm A (全 flag OFF、従来動作): Lego 12.5k full run repeat 3 (n=5, radix ballot default 上)
p3-tier1-bit-exactm2gpu-gatelegot1v2n5-regate
partial36.10 dB2m 8s412,0102026-08-30
t1v2-12500-off-r4M2 Tier1 再 gate v2 arm A (全 flag OFF、従来動作): Lego 12.5k full run repeat 4 (n=5, radix ballot default 上)
p3-tier1-bit-exactm2gpu-gatelegot1v2n5-regate
partial36.24 dB2m 7s407,1052026-08-30
t1v2-12500-off-r5M2 Tier1 再 gate v2 arm A (全 flag OFF、従来動作): Lego 12.5k full run repeat 5 (n=5, radix ballot default 上)
p3-tier1-bit-exactm2gpu-gatelegot1v2n5-regate
partial36.18 dB2m 6s407,8462026-08-30
t1v2-12500-off-r6M2 Tier1 再 gate v2 arm A (全 flag OFF、従来動作): Lego 12.5k full run repeat 1 (n=5, radix ballot default 上)
p3-tier1-bit-exactm2gpu-gatelegot1v2n5-regate
partial36.11 dB2m 7s406,2472026-08-30
t1v2-12500-on-r1M2 Tier1 再 gate v2 arm B (全 flag ON、A1-A4 候補): Lego 12.5k full run repeat 1 (n=5, radix ballot default 上)
p3-tier1-bit-exactm2gpu-gatelegot1v2n5-regate
partial36.14 dB2m 13s407,5672026-08-30
t1v2-12500-on-r2M2 Tier1 再 gate v2 arm B (全 flag ON、A1-A4 候補): Lego 12.5k full run repeat 2 (n=5, radix ballot default 上)
p3-tier1-bit-exactm2gpu-gatelegot1v2n5-regate
partial36.16 dB2m 5s409,5532026-08-30
t1v2-12500-on-r3M2 Tier1 再 gate v2 arm B (全 flag ON、A1-A4 候補): Lego 12.5k full run repeat 3 (n=5, radix ballot default 上)
p3-tier1-bit-exactm2gpu-gatelegot1v2n5-regate
partial36.20 dB2m 5s406,7142026-08-30
t1v2-12500-on-r4M2 Tier1 再 gate v2 arm B (全 flag ON、A1-A4 候補): Lego 12.5k full run repeat 4 (n=5, radix ballot default 上)
p3-tier1-bit-exactm2gpu-gatelegot1v2n5-regate
partial36.22 dB2m 5s405,4442026-08-30
t1v2-12500-on-r5M2 Tier1 再 gate v2 arm B (全 flag ON、A1-A4 候補): Lego 12.5k full run repeat 5 (n=5, radix ballot default 上)
p3-tier1-bit-exactm2gpu-gatelegot1v2n5-regate
partial36.20 dB2m 5s407,6132026-08-30
t1v2-12500-on-r6M2 Tier1 再 gate v2 arm B (全 flag ON、A1-A4 候補): Lego 12.5k full run repeat 1 (n=5, radix ballot default 上)
p3-tier1-bit-exactm2gpu-gatelegot1v2n5-regate
partial36.22 dB2m 8s406,1102026-08-30
final-parity-12500-r1msplat parity (最終スタック) — Lego 12500 iter、profile=false
p2-msplat-parity-finalmsplat-parityfinal-stacklego-12500
partial36.21 dB2m 33s407,5012026-08-17
final-parity-12500-r2msplat parity (最終スタック) — Lego 12500 iter、profile=false
p2-msplat-parity-finalmsplat-parityfinal-stacklego-12500
partial36.14 dB2m 32s408,1112026-08-17
final-parity-12500-r3msplat parity (最終スタック) — Lego 12500 iter、profile=false
p2-msplat-parity-finalmsplat-parityfinal-stacklego-12500
partial36.19 dB2m 44s411,1922026-08-17
final-parity-12500-r4msplat parity (最終スタック) — Lego 12500 iter、profile=false
p2-msplat-parity-finalmsplat-parityfinal-stacklego-12500
partial36.21 dB2m 48s408,5212026-08-17
final-parity-12500-r5msplat parity (最終スタック) — Lego 12500 iter、profile=false
p2-msplat-parity-finalmsplat-parityfinal-stacklego-12500
partial36.13 dB2m 42s405,1732026-08-17
final-parity-13300msplat parity (最終スタック) — Lego 13300 iter、profile=false
p2-msplat-parity-finalmsplat-parityfinal-stacklego-13300
partial36.15 dB2m 45s414,2052026-08-17
final-parity-30000msplat parity (最終スタック) — Lego 30000 iter、profile=false
p2-msplat-parity-finalmsplat-parityfinal-stacklego-30000
partial36.44 dB8m 18s468,5272026-08-17
final-parity-30000-r2msplat parity (最終スタック) — Lego 30000 iter、profile=false
p2-msplat-parity-finalmsplat-parityfinal-stacklego-30000
partial36.46 dB7m 59s467,7522026-08-17
final-parity-6000msplat parity (最終スタック) — Lego 6000 iter、profile=false
p2-msplat-parity-finalmsplat-parityfinal-stacklego-6000
partial34.89 dB51.60s319,4522026-08-17
counter-12500-noprof-r1counter session step 4: msplat との GPU 電力 interleave 比較用の本実装 12.5k (profile=false)。同時に当日の S2 gate baseline wallclock も兼ねる
p2-counter-sessionpowermsplat-interleaveno-profilelego-12500
partial36.20 dB3m 26s408,2432026-08-16
counter-12500-noprof-r2counter session step 4: msplat との GPU 電力 interleave 比較用の本実装 12.5k (profile=false)。同時に当日の S2 gate baseline wallclock も兼ねる
p2-counter-sessionpowermsplat-interleaveno-profilelego-12500
partial36.13 dB3m 39s407,9122026-08-16
counter-12500-profcounter session (H1/H2/H3): default スタック 12.5k を profile=true で 1 本 — [gpu] busy 軌跡 (refine tick / 序盤大 splat 期 vs 後期) を取る。wallclock は profiling オーバーヘッド込みなので 219.51s 系列とは比較不可
p2-counter-sessiongpu-busyprofile-onlego-12500
partial36.26 dB3m 56s409,1402026-08-16
lego-bench-calib-checkpointsP2 bench harness calibration — 3-regime (iter1000/5000/final) checkpoint 生成用 (lego-gap-baseline-12500 と同一 recipe、save_ply_at 追加のみ)
p1-gphase-gearly-stopsh-progressivestackedlego-15kbrush-compatpremultiplied
partial36.23 dB6m 0s409,6602026-08-16
lego-c0-ab-periter-1kC0 A/B smoke (1k) — arm A: binning = per_iter_alloc (C1 radix pass skip は両 arm 共通)
p2-gap-sweepc0-resident-arenaab-smokeper-iter-alloclego
partial19.48 dB11.36s6,8492026-08-16
lego-c0-ab-resident-1kC0 A/B smoke (1k) — arm B: binning = resident (C0 arena + C1)
p2-gap-sweepc0-resident-arenaab-smokeresidentlego
partial19.44 dB10.66s6,7912026-08-16
lego-c0-day0-attrib-1kC0 Day 0 gate — emit_pairs の host alloc/upload/readback attribution 実測 (1k smoke)
p2-gap-sweepc0-resident-arenaday0-gateattributionsmokelego
partial19.48 dB14.27s6,7882026-08-16
lego-c0-resident-125002026-08 gap sweep Phase 2 C0+C1 gate — baseline 12.5k に binning = resident を追加 (bit-exact 経路)
p2-gap-sweepc0-resident-arenac1-radix-pass-skipgate-runlego-12500
partial36.18 dB6m 35s407,9012026-08-16
lego-c3-tile-counting-12500C3 (per-tile counting sort) GPU gate — full 12.5k quality guard run vs lego-bench-calib-checkpoints baseline (PSNR 36.226 dB / 360s)
p1-gphase-gearly-stopsh-progressivestackedlego-15kbrush-compatpremultiplied
partial36.19 dB5m 29s405,2932026-08-16
lego-gap-baseline-125002026-08 msplat gap sweep Phase 0.1 — 規約 v2 適用後 baseline、per-kernel breakdown 取得
p1-gphase-gearly-stopsh-progressivestackedlego-15kbrush-compatpremultiplied
partial36.22 dB12m 46s407,2672026-08-16
lego-lc-smoke-1k2026-08 gap sweep Phase 2 2-BWD — raster_backward_range = last_contributor の 1k smoke
p2-gap-sweep2-bwd-last-contributorsmokelego-1k
partial19.57 dB9.94s6,8692026-08-16
lego-p2lc-125002026-08 gap sweep Phase 2 2-BWD gate — raster_backward_range = last_contributor 単独 (vs 395.59s baseline)
p2-gap-sweep2-bwd-last-contributorgate-runlego-12500
partial36.19 dB5m 36s408,7522026-08-16
lego-parity-noprof-12500msplat parity 比較 — Phase 2 default スタック (resident + last_contributor) 12.5k、profile=false の wallclock 実測
p2-gap-sweepmsplat-parityno-profilelego-12500
partial36.14 dB6m 25s412,6882026-08-16
lego-parity-noprof-7500msplat parity 比較 — Phase 2 default スタック (resident + last_contributor) 7.5k (Phase H pareto の 35.51 dB 点)、profile=false の wallclock 実測
p2-gap-sweepmsplat-parityno-profilelego-7500
partial35.39 dB3m 1s346,8602026-08-16
lego-s1-resident-125002026-08 msplat gap sweep S1 — pipeline 全域 resident 化 (Tier A/B/C) の 12.5k PSNR guard + wallclock gate
p2-s1residentalloc-auditgpu-gatelego-12500brush-compatpremultiplied
partial36.05 dB4m 32s407,5562026-08-16
lego-s1-resident-12500-rep22026-08 msplat gap sweep S1 — pipeline 全域 resident 化 (Tier A/B/C) の 12.5k PSNR guard + wallclock gate (rep2 — rep1 PSNR 36.051 が guard 外のため n=2 で分散確認)
p2-s1residentalloc-auditgpu-gatelego-12500brush-compatpremultiplied
partial36.17 dB4m 17s406,3212026-08-16
lego-s1-resident-12500-rep32026-08 msplat gap sweep S1 — pipeline 全域 resident 化 (Tier A/B/C) の 12.5k PSNR guard + wallclock gate (rep3 — rep1 36.051 guard 外 / rep2 36.168 guard 内、n=3 で分散確定)
p2-s1residentalloc-auditgpu-gatelego-12500brush-compatpremultiplied
partial36.22 dB4m 14s406,3042026-08-16
lego-stack-lc-stp-125002026-08 gap sweep Phase 2 最終スタック — resident + last_contributor + stop_the_pop (vs 395.59s)
p2-gap-sweep2-bwd-last-contributorfeat-f-stop-the-popstacklego-12500
partial36.14 dB5m 43s409,1412026-08-16
lego-stp-125002026-08 msplat gap sweep Phase 0.1 — 規約 v2 適用後 baseline、per-kernel breakdown 取得 (StopThePop 12.5k gate run)
p1-gphase-gearly-stopsh-progressivestackedlego-15kbrush-compatpremultiplied
partial36.19 dB11m 48s407,5752026-08-16
lego-stp-ab-original-5k2026-08 gap sweep Phase 1a — StopThePop A/B (original) 5k
p1-gphase-gearly-stopsh-progressivestackedlego-15kbrush-compatpremultiplied
partial34.59 dB3m 7s300,9872026-08-16
lego-stp-ab-original-5k2026-08 gap sweep Phase 1a — StopThePop A/B (original) 5k
p1-gphase-gearly-stopsh-progressivestackedlego-15kbrush-compatpremultiplied
partial34.10 dB3m 4s307,8012026-08-16
lego-stp-ab-stp-5k2026-08 gap sweep Phase 1a — StopThePop A/B (stp) 5k
p1-gphase-gearly-stopsh-progressivestackedlego-15kbrush-compatpremultiplied
partial34.18 dB3m 0s305,6832026-08-16
lego-stp-ab-stp-5k2026-08 gap sweep Phase 1a — StopThePop A/B (stp) 5k
p1-gphase-gearly-stopsh-progressivestackedlego-15kbrush-compatpremultiplied
partial34.07 dB3m 11s310,3292026-08-16
lego-stp-c0-125002026-08 gap sweep Phase 2 STP 再判定 — C0 後 baseline で emit_pairs = stop_the_pop 単独 (vs 395.59s)
p2-gap-sweepfeat-f-stop-the-popre-evalgate-runlego-12500
partial36.18 dB6m 25s408,0722026-08-16
lego-w1-warpany-125002026-08 1× 工程表 v2 W1 — warp sparse skip + exp-skip (bit-exact) の 12.5k PSNR guard + wallclock gate (rep1/3 — PSNR guard は n=3 mean 判定)
p2-w1warp-sparse-skipexp-skipgpu-gatelego-12500brush-compatpremultiplied
partial36.21 dB3m 59s410,4462026-08-16
lego-w1-warpany-12500-rep22026-08 1× 工程表 v2 W1 — warp sparse skip + exp-skip (bit-exact) の 12.5k PSNR guard + wallclock gate (rep2/3 — PSNR guard は n=3 mean 判定)
p2-w1warp-sparse-skipexp-skipgpu-gatelego-12500brush-compatpremultiplied
partial36.20 dB4m 6s406,7702026-08-16
lego-w1-warpany-12500-rep32026-08 1× 工程表 v2 W1 — warp sparse skip + exp-skip (bit-exact) の 12.5k PSNR guard + wallclock gate (rep3/3 — PSNR guard は n=3 mean 判定)
p2-w1warp-sparse-skipexp-skipgpu-gatelego-12500brush-compatpremultiplied
partial36.19 dB3m 57s411,2382026-08-16
lego-w2-chained-125002026-08 1× 工程表 v2 W2 — encoder/CB 統合 chained submission (bit-exact) の 12.5k PSNR guard + wallclock gate (rep1/3 — PSNR guard は n=3 mean 判定)
p2-w2chained-submissionencoder-cbgpu-gatelego-12500brush-compatpremultiplied
partial36.16 dB3m 33s408,2582026-08-16
lego-w2-chained-12500-rep22026-08 1× 工程表 v2 W2 — encoder/CB 統合 chained submission (bit-exact) の 12.5k PSNR guard + wallclock gate (rep2/3 — PSNR guard は n=3 mean 判定)
p2-w2chained-submissionencoder-cbgpu-gatelego-12500brush-compatpremultiplied
partial36.20 dB3m 44s407,1352026-08-16
lego-w2-chained-12500-rep32026-08 1× 工程表 v2 W2 — encoder/CB 統合 chained submission (bit-exact) の 12.5k PSNR guard + wallclock gate (rep3/3 — PSNR guard は n=3 mean 判定)
p2-w2chained-submissionencoder-cbgpu-gatelego-12500brush-compatpremultiplied
partial36.15 dB3m 40s413,2232026-08-16
lego-w2-smoke-1k-chainedW2 chained submission 1k smoke (submission = chained, scale_reg weight = 0.0)
p2-w2chained-submissionsmokelego-1k
partial19.60 dB7.01s6,8562026-08-16
lego-w2-smoke-1k-perkernelW2 chained submission 1k smoke (submission = per_kernel, scale_reg weight = 0.0)
p2-w2chained-submissionsmokelego-1k
partial19.61 dB7.85s6,8312026-08-16
lego-w2-smoke-1k-rmw-chainedW2 chained submission 1k smoke (submission = chained, scale_reg weight = 0.001)
p2-w2chained-submissionsmokelego-1k
partial18.08 dB9.10s6,2692026-08-16
lego-w2-smoke-1k-rmw-perkernelW2 chained submission 1k smoke (submission = per_kernel, scale_reg weight = 0.001)
p2-w2chained-submissionsmokelego-1k
partial18.06 dB9.98s6,2062026-08-16
refine-nondrain-smoke-1000-gpu-snapshotC2 施策 #4 smoke: refine_gpu_snapshot=true、crash なし + BackendConfig line 確認 (wallclock 判定はしない — 1k regime では refine tick 数・splat 数が gate 規模に遠く及ばない)
p3-refine-nondrainc2smoke
partial19.55 dB3.77s6,8162026-08-16
refine-nondrain-smoke-1000-offC2 施策 #4 smoke: refine_gpu_snapshot=true、crash なし + BackendConfig line 確認 (wallclock 判定はしない — 1k regime では refine tick 数・splat 数が gate 規模に遠く及ばない)
p3-refine-nondrainc2smoke
partial19.55 dB4.25s6,9482026-08-16
revisit-losspath-gpu-12500-r1保持 flag 再 gate sweep: loss_path = gpu の 12.5k full run repeat 1 (同一 session 交互 A/B)
p2-flag-revisit-sweeploss-pathgpu-gateno-profilelego-12500gpu
partial36.11 dB2m 59s405,9112026-08-16
revisit-losspath-gpu-12500-r2保持 flag 再 gate sweep: loss_path = gpu の 12.5k full run repeat 2 (同一 session 交互 A/B)
p2-flag-revisit-sweeploss-pathgpu-gateno-profilelego-12500gpu
partial36.24 dB3m 6s407,8222026-08-16
revisit-losspath-gpu-12500-r3保持 flag 再 gate sweep: loss_path = gpu の 12.5k full run repeat 3 (同一 session 交互 A/B)
p2-flag-revisit-sweeploss-pathgpu-gateno-profilelego-12500gpu
partial36.21 dB3m 4s406,7232026-08-16
revisit-losspath-host-12500-r1保持 flag 再 gate sweep: loss_path = host の 12.5k full run repeat 1 (同一 session 交互 A/B)
p2-flag-revisit-sweeploss-pathgpu-gateno-profilelego-12500host
partial36.18 dB4m 25s407,1812026-08-16
revisit-losspath-host-12500-r2保持 flag 再 gate sweep: loss_path = host の 12.5k full run repeat 2 (同一 session 交互 A/B)
p2-flag-revisit-sweeploss-pathgpu-gateno-profilelego-12500host
partial36.10 dB4m 22s409,8642026-08-16
revisit-losspath-host-12500-r3保持 flag 再 gate sweep: loss_path = host の 12.5k full run repeat 3 (同一 session 交互 A/B)
p2-flag-revisit-sweeploss-pathgpu-gateno-profilelego-12500host
partial36.15 dB5m 22s411,2702026-08-16
s2-12500-noprof-r1S2 gate: forward 1 CB 化 (chained forward + hierarchical GPU scan) の 12.5k full run repeat 1
p2-s2-forward-chaineds2gpu-gateno-profilelego-12500
partial36.15 dB3m 11s407,7512026-08-16
s2-12500-noprof-r2S2 gate: forward 1 CB 化 (chained forward + hierarchical GPU scan) の 12.5k full run repeat 2
p2-s2-forward-chaineds2gpu-gateno-profilelego-12500
partial36.19 dB3m 1s410,5022026-08-16
s2-12500-noprof-r3S2 gate: forward 1 CB 化 (chained forward + hierarchical GPU scan) の 12.5k full run repeat 3
p2-s2-forward-chaineds2gpu-gateno-profilelego-12500
partial36.15 dB3m 1s409,8612026-08-16
s2gate-base-12500-r1S2 gate 同一 session baseline: default スタック 12.5k repeat 1 (S2 arm と交互実行)
p2-s2-forward-chaineds2gpu-gatebaselineno-profilelego-12500
partial36.11 dB3m 28s405,2672026-08-16
s2gate-base-12500-r2S2 gate 同一 session baseline: default スタック 12.5k repeat 2 (S2 arm と交互実行)
p2-s2-forward-chaineds2gpu-gatebaselineno-profilelego-12500
partial36.17 dB3m 24s408,1232026-08-16
s2gate-base-12500-r3S2 gate 同一 session baseline: default スタック 12.5k repeat 3 (S2 arm と交互実行)
p2-s2-forward-chaineds2gpu-gatebaselineno-profilelego-12500
partial36.20 dB3m 22s404,0022026-08-16
s3-12500-depth2-profS3 gate arm B (pipelining = depth2): Lego 12.5k full run (profile=true、Layer 1 busy% / sync-tax 取得用)
p2-s3-iteration-pipelinings3gpu-gatelego
partial36.25 dB2m 42s410,0112026-08-16
s3-12500-depth2-r1S3 gate arm B (pipelining = depth2): Lego 12.5k full run repeat 1
p2-s3-iteration-pipelinings3gpu-gatelego
partial36.20 dB2m 29s405,0842026-08-16
s3-12500-depth2-r2S3 gate arm B (pipelining = depth2): Lego 12.5k full run repeat 2
p2-s3-iteration-pipelinings3gpu-gatelego
partial36.18 dB2m 32s407,6772026-08-16
s3-12500-depth2-r3S3 gate arm B (pipelining = depth2): Lego 12.5k full run repeat 3
p2-s3-iteration-pipelinings3gpu-gatelego
partial36.23 dB2m 32s406,5552026-08-16
s3-12500-off-profS3 gate arm A (pipelining = off): Lego 12.5k full run (profile=true、Layer 1 busy% / sync-tax 取得用)
p2-s3-iteration-pipelinings3gpu-gatelego
partial36.13 dB3m 0s407,4892026-08-16
s3-12500-off-r1S3 gate arm A (pipelining = off): Lego 12.5k full run repeat 1
p2-s3-iteration-pipelinings3gpu-gatelego
partial36.18 dB2m 59s406,4002026-08-16
s3-12500-off-r2S3 gate arm A (pipelining = off): Lego 12.5k full run repeat 2
p2-s3-iteration-pipelinings3gpu-gatelego
partial36.19 dB3m 11s409,0112026-08-16
s3-12500-off-r3S3 gate arm A (pipelining = off): Lego 12.5k full run repeat 3
p2-s3-iteration-pipelinings3gpu-gatelego
partial36.19 dB3m 3s411,2112026-08-16
s3-smoke-1000-depth2S3 smoke arm B (pipelining = depth2): Lego 1k、profile=true で busy% / sync-tax を取る
p2-s3-iteration-pipelinings3gpu-gatelego
partial19.60 dB4.62s6,8262026-08-16
s3-smoke-1000-offS3 smoke arm A (pipelining = off): Lego 1k、profile=true で busy% / sync-tax を取る
p2-s3-iteration-pipelinings3gpu-gatelego
partial19.59 dB4.69s6,7972026-08-16
bicycle-mipnerf-30kG1 Mip-NeRF 360 実写 3-way Pareto の 1 point — bicycle 30k full run
mipnerf360colmapbicycle30kpareto
partial0.00 dB1h 49m 9s311,2592026-07-22
bicycle-mipnerf-smoke-500G1 COLMAP loader smoke — bicycle 500 iter で loader + camera + SfM init 動作確認
mipnerf360colmaploadersmokebicycle
partial0.00 dB32.44s54,2882026-07-22
bicycle-mipnerf-tuned-3mG1 Mip-NeRF 360 実写 Pareto — bicycle 30k tuned recipe (refine.stop_iter=25000, capacity=3M)
mipnerf360colmapbicycle30kparetotuned
partial0.00 dB1h 42m 16s483,0192026-07-22
lego-mcmc-fix3-ablation-A-baseline-iter2000G3 MCMC fix3 smoke #1 — ablation A (all-MCMC OFF baseline) (fix2 の 1/10) で 2000 iter で PSNR 20 dB 到達を判定
mcmcg3fix3smokenoise-sweeplego
partial15.23 dB1m 2s149,7252026-07-01
lego-mcmc-fix3-ablation-B-scalel1only-iter2000G3 MCMC fix3 smoke #1 — ablation B (scale_l1=0.001 only) (fix2 の 1/10) で 2000 iter で PSNR 20 dB 到達を判定
mcmcg3fix3smokenoise-sweeplego
partial10.84 dB46.49s63,3582026-07-01
lego-mcmc-fix3-ablation-C-relocationonly-iter2000G3 MCMC fix3 smoke #1 — ablation C (relocation only, noise ≈0) (fix2 の 1/10) で 2000 iter で PSNR 20 dB 到達を判定
mcmcg3fix3smokenoise-sweeplego
partial15.68 dB1m 14s205,6132026-07-01
lego-mcmc-fix3-final-noise-relocation-30kG3 MCMC final — noise 5e2 + relocation (scale_l1=0)、fix3 犯人除去版 で Lego 30k、Phase I との差分で MCMC 効果を測定
mcmcg3fix3smokenoise-sweeplego
partial14.68 dB2h 38m 2s1,840,6042026-07-01
lego-mcmc-fix3-final-relocation-only-30kG3 MCMC final — relocation only (scale_l1=0, noise=5e0)、ablation-C 30k full で Lego 30k、Phase I との差分で growth schedule 効果を測定
mcmcg3fix3smokenoise-sweeplego
partial14.65 dB2h 12m 29s1,315,6052026-07-01
lego-mcmc-fix3-noise5e1-iter2000G3 MCMC fix3 smoke #1 — noise 5e1 (fix2 の 1/100、calibration doc §案 A の下限側) で 2000 iter で PSNR 20 dB 到達を判定
mcmcg3fix3smokenoise-sweeplego
partial9.74 dB55.95s53,5872026-07-01
lego-mcmc-fix3-noise5e2-iter2000G3 MCMC fix3 smoke #1 — noise 5e2 (fix2 の 1/10) で 2000 iter で PSNR 20 dB 到達を判定
mcmcg3fix3smokenoise-sweeplego
partial10.03 dB1m 15s61,3742026-07-01
lego-mcmc-fix3-noise5e2-smokeG3 MCMC fix3 smoke #1 — noise 5e2 (fix2 の 1/10) で 500 iter loss trajectory 観察
mcmcg3fix3smokenoise-sweeplego
partial11.92 dB8.01s9,1232026-07-01
lego-mcmc-fix3-noise5e3-iter2000G3 MCMC fix3 smoke #1 — noise 5e3 で 2000 iter で PSNR 20 dB 到達を判定
mcmcg3fix3smokenoise-sweeplego
partial10.29 dB59.43s54,1812026-07-01
lego-mcmc-fix3-noise5e3-smokeG3 MCMC fix3 smoke #1 — noise 5e3 で 500 iter loss trajectory 観察
mcmcg3fix3smokenoise-sweeplego
partial11.97 dB7.16s10,0602026-07-01
chair-brushcompat-opacdecay-30kP1.D multi-scene re-chain: chair brushcompat + opacity_decay 30k
p1-dmulti-scenebrush-compatpremultipliedopacity-decayphase-d-multichair
partial35.81 dB58m 58s800,6452026-05-25
chair-phase-g1-stop15kP1.D multi-scene re-chain: chair brushcompat + opacity_decay 30k
p1-gphase-gearly-stopbrush-compatpremultiplied
partial35.90 dB22m 19s653,4732026-05-25
chair-phase-g1g3-stacked-15kP1.D multi-scene re-chain: chair brushcompat + opacity_decay 30k
p1-gphase-gearly-stopsh-progressivestackedbrush-compatpremultiplied
partial35.95 dB24m 58s862,8252026-05-25
chair-phase-g3-sh-progressive-30kP1.D multi-scene re-chain: chair brushcompat + opacity_decay 30k
p1-dmulti-scenebrush-compatpremultipliedopacity-decayphase-d-multichair
partial35.95 dB1h 15m 42s1,148,6672026-05-25
chair-phase-i-adaptive-12500Phase I scene-adaptive: chair @ stop_iter=12500 (fast converger 仮説検証) variant (early stop + SH progressive)
p1-gphase-gearly-stopsh-progressivestackedlego-15kbrush-compatpremultiplied
partial35.88 dB18m 35s794,5622026-05-25
drums-brushcompat-opacdecay-30kP1.D multi-scene re-chain: drums brushcompat + opacity_decay 30k
p1-dmulti-scenebrush-compatpremultipliedopacity-decayphase-d-multidrums
partial27.20 dB1h 1m 50s868,3042026-05-25
drums-phase-g1-stop15kP1.D multi-scene re-chain: drums brushcompat + opacity_decay 30k
p1-gphase-gearly-stopbrush-compatpremultiplied
partial26.76 dB22m 33s786,1902026-05-25
drums-phase-g1g3-stacked-15kP1.D multi-scene re-chain: drums brushcompat + opacity_decay 30k
p1-gphase-gearly-stopsh-progressivestackedbrush-compatpremultiplied
partial25.38 dB22m 32s868,1992026-05-25
drums-phase-g3-sh-progressive-30kP1.D multi-scene re-chain: drums brushcompat + opacity_decay 30k
p1-dmulti-scenebrush-compatpremultipliedopacity-decayphase-d-multidrums
partial27.22 dB1h 5m 36s1,001,0142026-05-25
ficus-brushcompat-opacdecay-30kP1.D multi-scene re-chain: ficus brushcompat + opacity_decay 30k
p1-dmulti-scenebrush-compatpremultipliedopacity-decayphase-d-multificus
partial34.22 dB21m 38s203,4122026-05-25
ficus-phase-g1-stop15kP1.D multi-scene re-chain: ficus brushcompat + opacity_decay 30k
p1-gphase-gearly-stopbrush-compatpremultiplied
partial32.43 dB10m 31s244,7932026-05-25
ficus-phase-g1g3-stacked-15kP1.D multi-scene re-chain: ficus brushcompat + opacity_decay 30k
p1-gphase-gearly-stopsh-progressivestackedbrush-compatpremultiplied
partial31.07 dB10m 25s267,6182026-05-25
ficus-phase-g3-sh-progressive-30kP1.D multi-scene re-chain: ficus brushcompat + opacity_decay 30k
p1-dmulti-scenebrush-compatpremultipliedopacity-decayphase-d-multificus
partial34.28 dB21m 40s226,7492026-05-25
ficus-phase-h-validate-12500Phase H validation: ficus @ stop_iter=12500 (Lego sweet spot が scene-dependent か検証) variant (early stop + SH progressive)
p1-gphase-gearly-stopsh-progressivestackedlego-15kbrush-compatpremultiplied
partial30.10 dB8m 29s274,7372026-05-25
hotdog-brushcompat-opacdecay-30kP1.D multi-scene re-chain: hotdog brushcompat + opacity_decay 30k
p1-dmulti-scenebrush-compatpremultipliedopacity-decayphase-d-multihotdog
partial37.33 dB23m 39s206,6202026-05-25
hotdog-phase-g1-stop15kP1.D multi-scene re-chain: hotdog brushcompat + opacity_decay 30k
p1-gphase-gearly-stopbrush-compatpremultiplied
partial35.88 dB9m 30s184,4062026-05-25
hotdog-phase-g1g3-stacked-15kP1.D multi-scene re-chain: hotdog brushcompat + opacity_decay 30k
p1-gphase-gearly-stopsh-progressivestackedbrush-compatpremultiplied
partial37.19 dB10m 42s262,8292026-05-25
hotdog-phase-g3-sh-progressive-30kP1.D multi-scene re-chain: hotdog brushcompat + opacity_decay 30k
p1-dmulti-scenebrush-compatpremultipliedopacity-decayphase-d-multihotdog
partial37.37 dB28m 17s310,0452026-05-25
lego-brushcompat-opacdecay-r0p001-5kP1.D rate sweep r=0.001 (vs default 0.004)
p1-drate-sweeplego-5kbrush-compatpremultipliedopacity-decay
partial31.95 dB2m 26s88,6852026-05-25
lego-brushcompat-opacdecay-r0p002-5kP1.D rate sweep r=0.002 (vs default 0.004)
p1-drate-sweeplego-5kbrush-compatpremultipliedopacity-decay
partial32.09 dB3m 1s87,9402026-05-25
lego-brushcompat-opacdecay-r0p006-5kP1.D rate sweep r=0.006 (vs default 0.004)
p1-drate-sweeplego-5kbrush-compatpremultipliedopacity-decay
partial31.65 dB3m 24s81,3272026-05-25
lego-brushcompat-opacdecay-r0p008-5kP1.D rate sweep r=0.008 (vs default 0.004)
p1-drate-sweeplego-5kbrush-compatpremultipliedopacity-decay
partial31.81 dB2m 29s80,2062026-05-25
lego-phase-f1-baseline-5kPhase F.1 baseline (use_simd_emit=false) re-take for A/B
p1-profileaxis-1phase-fab-baselinelego-5kpremultiplied
partial31.63 dB1m 52s82,3382026-05-25
lego-phase-f1-emit-simd-5kPhase F.1 emit_pairs_simd gate flip 5k smoke (clean baseline 比較用)
p1-profileaxis-1phase-femit-simdlego-5kpremultipliedtier-1
partial31.59 dB1m 57s84,2082026-05-25
lego-phase-f2-f16-fwd-5kPhase F.2 f16 forward kernel A/B test 5k smoke
p1-profileaxis-1phase-ff16-forwardlego-5kpremultipliedtier-1
partial31.80 dB1m 54s87,1812026-05-25
lego-phase-f3-baseline-5kPhase F.3 baseline (SPLAT_RADIX_CPU_SCAN=1, legacy CPU scan) for radix GPU scan A/B
p1-profileaxis-1phase-ff3-baselinelego-5kpremultipliedcpu-scan
partial31.60 dB1m 55s84,8402026-05-25
lego-phase-f3-gpu-scan-5kPhase F.3 GPU prefix scan (default) — radix_sort with intra-pass waits removed
p1-profileaxis-1phase-ff3-gpu-scanlego-5kpremultipliedradix-gpu-scan
partial31.87 dB1m 58s84,3652026-05-25
lego-phase-g1-stop15kPhase G.1 early stop @ iter 15000 (Lego validation 先行)
p1-gphase-gearly-stoplego-15kbrush-compatpremultiplied
partial35.69 dB14m 30s335,0342026-05-25
lego-phase-g1g3-stacked-15kPhase G.1+G.3 stacked variant (early stop + SH progressive)
p1-gphase-gearly-stopsh-progressivestackedlego-15kbrush-compatpremultiplied
partial36.25 dB16m 12s427,8822026-05-25
lego-phase-g3-sh-progressive-30kPhase G.3 SH progressive 30k full validation
p1-gphase-gsh-progressivelego-30kbrush-compatpremultiplied
partial36.38 dB41m 7s487,7412026-05-25
lego-phase-g3-sh-progressive-5kPhase G.3 SH degree progressive growth (0→3, unlock_interval=1000) Lego 5k smoke
p1-profileaxis-1phase-gsh-progressivelego-5kpremultiplied
partial31.51 dB1m 36s64,3302026-05-25
lego-phase-h-pareto-10000Phase H Pareto sweep @ stop_iter=10000 variant (early stop + SH progressive)
p1-gphase-gearly-stopsh-progressivestackedlego-15kbrush-compatpremultiplied
partial35.93 dB8m 42s385,0072026-05-25
lego-phase-h-pareto-12500Phase H Pareto sweep @ stop_iter=12500 variant (early stop + SH progressive)
p1-gphase-gearly-stopsh-progressivestackedlego-15kbrush-compatpremultiplied
partial36.26 dB11m 48s407,0532026-05-25
lego-phase-h-pareto-17500Phase H Pareto sweep @ stop_iter=17500 variant (early stop + SH progressive)
p1-gphase-gearly-stopsh-progressivestackedlego-15kbrush-compatpremultiplied
partial36.35 dB18m 56s438,2382026-05-25
lego-phase-h-pareto-20000Phase H Pareto sweep @ stop_iter=20000 variant (early stop + SH progressive)
p1-gphase-gearly-stopsh-progressivestackedlego-15kbrush-compatpremultiplied
partial36.36 dB22m 40s448,4012026-05-25
lego-phase-h-pareto-25000Phase H Pareto sweep @ stop_iter=25000 variant (early stop + SH progressive)
p1-gphase-gearly-stopsh-progressivestackedlego-15kbrush-compatpremultiplied
partial36.09 dB30m 20s464,2352026-05-25
lego-phase-h-pareto-7500Phase H Pareto sweep @ stop_iter=7500 variant (early stop + SH progressive)
p1-gphase-gearly-stopsh-progressivestackedlego-15kbrush-compatpremultiplied
partial35.51 dB6m 20s347,4892026-05-25
lego-profile-clean-5kClean single-process profile baseline 取り直し (contention 解消後)
p1-profileaxis-1clean-baselinelego-5kpremultipliedsingle-process
partial31.51 dB1m 55s84,8312026-05-25
materials-brushcompat-opacdecay-30kP1.D multi-scene re-chain: materials brushcompat + opacity_decay 30k
p1-dmulti-scenebrush-compatpremultipliedopacity-decayphase-d-multimaterials
partial29.90 dB27m 27s266,0182026-05-25
materials-phase-g1-stop15kP1.D multi-scene re-chain: materials brushcompat + opacity_decay 30k
p1-gphase-gearly-stopbrush-compatpremultiplied
partial29.90 dB11m 28s263,6882026-05-25
materials-phase-g1g3-stacked-15kP1.D multi-scene re-chain: materials brushcompat + opacity_decay 30k
p1-gphase-gearly-stopsh-progressivestackedbrush-compatpremultiplied
partial29.99 dB12m 59s343,5082026-05-25
materials-phase-g3-sh-progressive-30kP1.D multi-scene re-chain: materials brushcompat + opacity_decay 30k
p1-dmulti-scenebrush-compatpremultipliedopacity-decayphase-d-multimaterials
partial30.03 dB30m 46s349,7842026-05-25
materials-phase-i-adaptive-12500Phase I scene-adaptive: materials @ stop_iter=12500 (fast converger 仮説検証) variant (early stop + SH progressive)
p1-gphase-gearly-stopsh-progressivestackedlego-15kbrush-compatpremultiplied
partial29.97 dB10m 59s340,7632026-05-25
mic-brushcompat-opacdecay-30kP1.D multi-scene re-chain: mic brushcompat + opacity_decay 30k
p1-dmulti-scenebrush-compatpremultipliedopacity-decayphase-d-multimic
partial36.38 dB32m 8s329,1102026-05-25
mic-phase-g1-stop15kP1.D multi-scene re-chain: mic brushcompat + opacity_decay 30k
p1-gphase-gearly-stopbrush-compatpremultiplied
partial30.54 dB11m 9s270,3712026-05-25
mic-phase-g1g3-stacked-15kP1.D multi-scene re-chain: mic brushcompat + opacity_decay 30k
p1-gphase-gearly-stopsh-progressivestackedbrush-compatpremultiplied
partial30.33 dB12m 14s330,5422026-05-25
mic-phase-g3-sh-progressive-30kP1.D multi-scene re-chain: mic brushcompat + opacity_decay 30k
p1-dmulti-scenebrush-compatpremultipliedopacity-decayphase-d-multimic
partial36.62 dB33m 29s391,3732026-05-25
mic-phase-h-validate-12500Phase H validation: mic @ stop_iter=12500 (Lego sweet spot が scene-dependent か検証) variant (early stop + SH progressive)
p1-gphase-gearly-stopsh-progressivestackedlego-15kbrush-compatpremultiplied
partial30.35 dB9m 18s298,2912026-05-25
multiscene-drums-30k多 scene 退行チェック (drums, G.3-alone 30k recipe, n=1)
p1-dmulti-scenebrush-compatpremultipliedopacity-decayphase-d-multidrums
partial27.21 dB11m 27s997,9582026-05-25
multiscene-ficus-30k多 scene 退行チェック (ficus, G.3-alone 30k recipe, n=1)
p1-dmulti-scenebrush-compatpremultipliedopacity-decayphase-d-multificus
partial34.26 dB5m 49s227,4812026-05-25
multiscene-ficus-30k-q静穏再実行 (ficus, G.3-alone 30k recipe, n=1, load-controlled)
p1-dmulti-scenebrush-compatpremultipliedopacity-decayphase-d-multificus
partial32.84 dB4m 14s228,1472026-05-25
multiscene-ficus-30k-q2静穏再実行 (ficus, G.3-alone 30k recipe, n=1, load-controlled)
p1-dmulti-scenebrush-compatpremultipliedopacity-decayphase-d-multificus
partial34.28 dB4m 16s228,5952026-05-25
multiscene-materials-30k多 scene 退行チェック (materials, G.3-alone 30k recipe, n=1)
p1-dmulti-scenebrush-compatpremultipliedopacity-decayphase-d-multimaterials
partial30.00 dB7m 1s355,0332026-05-25
multiscene-materials-30k-q静穏再実行 (materials, G.3-alone 30k recipe, n=1, load-controlled)
p1-dmulti-scenebrush-compatpremultipliedopacity-decayphase-d-multimaterials
partial30.00 dB5m 19s354,9172026-05-25
ship-brushcompat-opacdecay-30kP1.D multi-scene re-chain: ship brushcompat + opacity_decay 30k
p1-dmulti-scenebrush-compatpremultipliedopacity-decayphase-d-multiship
partial30.93 dB36m 8s373,6812026-05-25
ship-phase-g1-stop15kP1.D multi-scene re-chain: ship brushcompat + opacity_decay 30k
p1-gphase-gearly-stopbrush-compatpremultiplied
partial29.73 dB14m 38s378,7602026-05-25
ship-phase-g1g3-stacked-15kP1.D multi-scene re-chain: ship brushcompat + opacity_decay 30k
p1-gphase-gearly-stopsh-progressivestackedbrush-compatpremultiplied
partial29.82 dB16m 8s477,7302026-05-25
ship-phase-g3-sh-progressive-30kP1.D multi-scene re-chain: ship brushcompat + opacity_decay 30k
p1-dmulti-scenebrush-compatpremultipliedopacity-decayphase-d-multiship
partial30.88 dB39m 46s495,1602026-05-25
chair-brushcompat-30kP1.B+F Stage 2 multi-scene: chair brushcompat 30k
p1-b-fmulti-scenebrush-compatconvention-bridgepremultipliedstage-2-multichair
partial35.57 dB2h 3m 21s1,989,3952026-05-24
drums-brushcompat-30kP1.B+F Stage 2 multi-scene: drums brushcompat 30k
p1-b-fmulti-scenebrush-compatconvention-bridgepremultipliedstage-2-multidrums
partial26.61 dB2h 24m 29s2,541,7612026-05-24
ficus-brushcompat-30kP1.B+F Stage 2 multi-scene: ficus brushcompat 30k
p1-b-fmulti-scenebrush-compatconvention-bridgepremultipliedstage-2-multificus
partial32.95 dB56m 53s879,2262026-05-24
hotdog-brushcompat-30kP1.B+F Stage 2 multi-scene: hotdog brushcompat 30k
p1-b-fmulti-scenebrush-compatconvention-bridgepremultipliedstage-2-multihotdog
partial37.14 dB38m 47s485,3572026-05-24
lego-brushcompat-base-30kP1.B.F Stage 2 brush 互換 30k — gt_convention=premultiplied、refine 期間 brush 流
p1-b-flego-30kbrush-compatconvention-bridgepremultipliedm3-gate
partial35.18 dB1h 2m 18s846,6892026-05-24
lego-brushcompat-opacdecay-30kP1.D Stage 2 brush 互換 + opacity decay 30k full bench
p1-dopacity-decaylego-30kbrush-compatpremultipliedstage-2splat-efficient
partial36.11 dB41m 54s375,1462026-05-24
lego-brushcompat-opacdecay-5kP1.D opacity decay 5k — brushcompat-base-5k に opacity_decay_rate=0.004 を追加
p1-dlego-5ksmokebrush-compatopacity-decaysplat-count-reduction
partial31.69 dB2m 34s83,0932026-05-24
materials-brushcompat-30kP1.B+F Stage 2 multi-scene: materials brushcompat 30k
p1-b-fmulti-scenebrush-compatconvention-bridgepremultipliedstage-2-multimaterials
partial28.49 dB1h 29m 9s1,229,7582026-05-24
mic-brushcompat-30kP1.B+F Stage 2 multi-scene: mic brushcompat 30k
p1-b-fmulti-scenebrush-compatconvention-bridgepremultipliedstage-2-multimic
partial36.19 dB1h 39m 36s1,484,9332026-05-24
ship-brushcompat-30kP1.B+F Stage 2 multi-scene: ship brushcompat 30k
p1-b-fmulti-scenebrush-compatconvention-bridgepremultipliedstage-2-multiship
partial30.37 dB1h 4m 21s897,2592026-05-24
chair-batched-30kA.7 × multi-scene: chair 30k with SPLAT_BATCHED_FORWARD=1
auto-benchchair-batched-30k
partial22.78 dB18m 39s39,4512026-05-23
drums-batched-30kA.7 × multi-scene: drums 30k with SPLAT_BATCHED_FORWARD=1
auto-benchdrums-batched-30k
partial17.90 dB20m 43s64,5752026-05-23
ficus-batched-30kA.7 × multi-scene: ficus 30k with SPLAT_BATCHED_FORWARD=1
auto-benchficus-batched-30k
partial14.05 dB22m 32s61,5992026-05-23
hotdog-batched-30kA.7 × multi-scene: hotdog 30k with SPLAT_BATCHED_FORWARD=1
auto-benchhotdog-batched-30k
partial29.46 dB22m 35s80,8412026-05-23
kahan-500smokeA.10 Kahan/Neumaier compiler-survival A/B (500 iter quick smoke)
kahansmokecompiler-checka-10
partial13.22 dB12.89s5,4972026-05-23
kahan-500smokeA.10 Kahan/Neumaier compiler-survival A/B (500 iter quick smoke)
kahansmokecompiler-checka-10
partial13.22 dB14.77s5,4972026-05-23
lego-a7-batched-30kA.7 batched cmd buffer — extract+rasterize.forward と backward chain を 1 cmd buffer 集約
a-7icb-batchinglego-30kphase-5
partial24.58 dB21m 47s82,8552026-05-23
lego-a9-f16-30kA.9 f16 forward (half3 per-pixel accumulator) — wallclock vs PSNR trade-off
a-9f16-forwardmixed-precisionlego-30kphase-5
partial14.87 dB40m 39s94,9002026-05-23
lego-cap100000-30kE.6 capacity scaling: lego sh3 30k capacity=100000 (モバイル制約下の最適 capacity 探索)
auto-benchlego-cap100000-30k
partial24.88 dB21m 22s81,3252026-05-23
lego-cap200000-30kE.6 capacity scaling: lego sh3 30k capacity=200000 (モバイル制約下の最適 capacity 探索)
auto-benchlego-cap200000-30k
partial24.60 dB21m 42s85,0212026-05-23
lego-cap50000-30kE.6 capacity scaling: lego sh3 30k capacity=50000 (モバイル制約下の最適 capacity 探索)
auto-benchlego-cap50000-30k
partial25.28 dB21m 35s82,5502026-05-23
lego-cap500000-30kE.6 capacity scaling: lego sh3 30k capacity=500000 (モバイル制約下の最適 capacity 探索)
auto-benchlego-cap500000-30k
partial25.12 dB21m 28s80,8832026-05-23
lego-iter10000E.5 iter scaling: lego sh3 10000 iter (収益逓減 plateau 探索)
auto-benchlego-iter10000
partial24.01 dB6m 40s75,5102026-05-23
lego-iter15000E.5 iter scaling: lego sh3 15000 iter (収益逓減 plateau 探索)
auto-benchlego-iter15000
partial23.93 dB10m 32s81,8542026-05-23
lego-iter20000E.5 iter scaling: lego sh3 20000 iter (収益逓減 plateau 探索)
auto-benchlego-iter20000
partial23.87 dB14m 8s80,7162026-05-23
lego-iter25000E.5 iter scaling: lego sh3 25000 iter (収益逓減 plateau 探索)
auto-benchlego-iter25000
partial24.62 dB18m 19s84,0862026-05-23
lego-mcmc-fix2-30kA.2 MCMC 完全実装 — fix#2 (opacity_l1=0、scale_l1 弱、reorder 済) で Lego 30k
mcmcphase-5lego-30ka2-validationfix2
partial17.38 dB1h 10m 14s1,000,0002026-05-23
lego-variance-trial1-30kA.10 variance baseline trial 1/3 (seed=42 固定で GPU 非決定性のみ観測)
auto-benchlego-variance-trial1-30k
partial24.86 dB22m 7s84,1492026-05-23
lego-variance-trial2-30kA.10 variance baseline trial 2/3 (seed=42 固定で GPU 非決定性のみ観測)
auto-benchlego-variance-trial2-30k
partial24.36 dB22m 0s83,0792026-05-23
lego-variance-trial3-30kA.10 variance baseline trial 3/3 (seed=42 固定で GPU 非決定性のみ観測)
auto-benchlego-variance-trial3-30k
partial25.24 dB22m 20s86,4232026-05-23
materials-30kA.4 NeRF Synthetic materials 30k baseline (sh=3)
auto-benchmaterials-30k
partial12.71 dB28m 27s161,9352026-05-23
mcmc-combo-iter-bisectMCMC segfault 切り分け #3: noise + L1、log_every=1 で iter localize
debugmcmcsmokeiter-bisect
partial5.19 dB0.57s5,2072026-05-23
mcmc-fix2-scale-onlyMCMC 修正案 #2: opacity_l1=0、scale_l1=0.001、noise=5e3
debugmcmcfix-attempt-2no-opacity-l1
partial10.94 dB15.84s26,1372026-05-23
mcmc-l1-only-smokeMCMC segfault 切り分け #1: scale_eigen_l1 + opacity_l1 のみ
debugmcmcsmokel1-only
partial1.08 dB4.82s5,2072026-05-23
mcmc-noise-sh3-smokeMCMC segfault 切り分け #2: SGLD noise 5e3 のみ、sh=3 cap=1M
debugmcmcsmokenoise-only-sh3
partial12.90 dB7.64s5,2072026-05-23
mic-30kA.4 NeRF Synthetic mic 30k baseline (sh=3)
auto-benchmic-30k
partial15.03 dB22m 13s31,9012026-05-23
ship-30kA.4 NeRF Synthetic ship 30k baseline (sh=3)
auto-benchship-30k
partial15.04 dB23m 58s107,4502026-05-23
chair-30kA.4 NeRF Synthetic 他シーン展開 — chair 30k baseline (sh=3)
scene-ablationchair-30kscene-chairphase-5
partial22.88 dB19m 51s38,9282026-05-22
drums-30kA.4 NeRF Synthetic 他シーン展開 — drums 30k baseline (sh=3)
scene-ablationdrums-30kscene-drumsphase-5
partial17.77 dB21m 23s64,5152026-05-22
ficus-30kA.4 NeRF Synthetic 他シーン展開 — ficus 30k baseline (sh=3)
scene-ablationficus-30kscene-ficusphase-5
partial13.96 dB22m 48s60,9382026-05-22
hotdog-30kA.4 NeRF Synthetic 他シーン展開 — hotdog 30k baseline (sh=3)
scene-ablationhotdog-30kscene-hotdogphase-5
partial30.29 dB23m 46s82,1542026-05-22
lego-res200-30kA.12 Resolution scaling 実験 — Lego 200x200 px
resolution-scalinglego-30kres-200phase-5
partial23.90 dB23m 41s80,1472026-05-22
lego-res400-30kA.12 Resolution scaling 実験 — Lego 400x400 px
resolution-scalinglego-30kres-400phase-5
partial25.48 dB22m 50s83,1772026-05-22
lego-res800-30kA.12 Resolution scaling 実験 — Lego 800x800 px
resolution-scalinglego-30kres-800phase-5
partial24.96 dB22m 38s81,9452026-05-22
lego-sh0-30kA.8 SH degree ablation — sh_degree=0 (DC only, no view-dependence)
sh-ablationlego-30ksh-0phase-5
partial16.50 dB26m 47s40,2882026-05-22
lego-sh1-30kA.8 SH degree ablation — sh_degree=1 (DC only, no view-dependence)
sh-ablationlego-30ksh-1phase-5
partial19.07 dB22m 38s64,1392026-05-22
lego-sh2-30kA.8 SH degree ablation — sh_degree=2 (DC only, no view-dependence)
sh-ablationlego-30ksh-2phase-5
partial23.63 dB22m 37s70,1552026-05-22
lego-sh3-30kA.8 SH degree ablation — sh_degree=3 (DC only, no view-dependence)
sh-ablationlego-30ksh-3phase-5
partial24.88 dB23m 13s83,7342026-05-22
mcmc-all-smokeA.2 MCMC 完全実装 — SGLD noise + λ_Σ + λ_o + 5%/tick growth + multinomial relocation (paper §3.2/3.4 準拠)
mcmcphase-5lego-30ka2-validation
partial2.52 dB1.17s5,2072026-05-22
mcmc-noise-only-smokeA.2 segfault 切り分け: SGLD noise のみ
debugmcmcsmoke
partial9.86 dB2.29s5,2072026-05-22
mcmc-scale-l1-smokeA.2 segfault 切り分け: SGLD noise のみ
debugmcmcsmoke
partial8.06 dB2.03s5,2072026-05-22
2026-04-30-2321-m3x-30k-migration-gate30,000 反復 @ Lego / 単発、新 workspace の動作確認を目的とした 30k 比較 run。
m3xphase-cmigration-gatelego-30k
partial24.84 dB23m 40s79,2392026-04-30
smoke-1000iterPhase C7 完了確認、1000 iter で全 path が走ること
smokephase-c7
partial14.09 dB22.80s14,0522026-04-30
lego-smoke-l1ssimL1+SSIM loss (λ=0.2、K=7) を新規実装した直後の smoke run。200 iter / 50 iter refine 周期で SSIM kernel・refine・eval が一通り完走することを確認する用途。
legosmokel1ssim200-itersanitylegacy
reject12.39 dB16.8 s13,7532026-04-25
lego-10k-no-resetrefine schedule を [500, 7000]・opacity_reset を 0 (無効) に設定し、reset の有無が PSNR に与える影響を切り分ける 10k 反復 run。
lego10kl1no-resetablationlegacy
partial20.40 dB7m 55s5,4172026-04-24
lego-30k-v3refine [500, 15000]・opacity_reset 3000 の長期 run。reset がトリガされるたびに splat が prune され、densification より prune が dominate し続け、最終 5,667 splats / 8.26 dB に終わった。
lego30kl1reset-3000rejectlegacy
reject8.27 dB38m 10s5,6672026-04-24
lego-30k-v4opacity_reset を切り refine 範囲を [500, 15000] に拡大した結果、init 10k → 最終 376,239 splats に膨張。training loss は 9.5e-3 まで下がるが val PSNR は 12.62 dB と overfitting / 表現破綻。
lego30kl1no-resetovergrowthlegacy
reject12.62 dB1h 15m 19s376,2392026-04-24
lego-30k-v5F config (refine 範囲 [500, 1500] / reset 無効 / lr_pos 1.6e-4) を 30,000 反復に延長した run。
lego30kl1no-resetf-configlegacy
partial21.34 dB30m 44s79,1382026-04-24
lego-tune-aF config (refine 範囲 [500, 1500] / reset 無効) を tune-A baseline として確立した 10k 反復 run。
lego10kl1tune-af-configlegacy
partial22.25 dB10m 30s88,1882026-04-24
lego-tune-a-l1ssimtune-A の refine スケジュール ([500, 1500], reset 無効) を維持したまま、損失関数を L1 → L1+SSIM (λ=0.2) に置換した 10k 反復 run。
lego10kl1ssimtune-ano-resetlegacy
partial22.42 dB12m 30s120,3282026-04-24
lego-tune-a-l1ssim-scaleregtune-A + L1+SSIM 構成に scale_reg_w = 0.01 を加えた 10k 反復 run。splat 数が capacity 上限 (1M) まで暴走、PSNR は 13.46 dB に大幅劣化。
lego10kl1ssimscaleregnegativelegacy
reject13.46 dB23m 56s1,053,0752026-04-24
© 2026 大竹律輝 — 松田研究室