[{"data":1,"prerenderedAt":350},["ShallowReactive",2],{"finding:p1-mipnerf360-pareto-v2":3,"finding-runs:p1-mipnerf360-pareto-v2":256,"finding-related:p1-mipnerf360-pareto-v2":273},{"meta":4,"impact":39,"sections":45},{"config_paths":5,"id":12,"title":13,"subtitle":14,"eyebrow":15,"date":16,"status":17,"category":18,"polarity":19,"axes":20,"tags":22,"task_code":31,"related_runs":32,"related_findings":35},[6,7,8,9,10,11],"splat\u002Fconfigs\u002F2026-07-22-bicycle-30k.toml","splat\u002Fconfigs\u002F2026-07-22-bicycle-smoke-500.toml","splat\u002Fconfigs\u002F2026-07-22-bicycle-tuned-3m.toml","splat\u002Fconfigs\u002F2026-07-22-garden-30k.toml","splat\u002Fconfigs\u002F2026-07-22-counter-30k.toml","splat\u002Fconfigs\u002F2026-07-22-room-30k.toml","p1-mipnerf360-pareto-v2","G1 Mip-NeRF 360 実写 Pareto v2 — bicycle before \u002F after recipe tuning","2026-07-22 の G1 4 scene sweep で bicycle baseline (Phase I recipe 流用、refine.stop_iter=15000, capacity=1.5M) が post-hoc COLMAP eval で \u003Cstrong>PSNR 15.549 dB (25 views)\u003C\u002Fstrong> と Kerbl 2023 paper 25.25 dB 比 \u003Cstrong>-9.7 dB\u003C\u002Fstrong> の大 gap を示した。残り 3 scene の同 recipe 走行は ROI 低と判断して中断、bicycle 1 scene に絞って \u003Cstrong>tuned recipe (refine.stop_iter=25000, capacity=3M)\u003C\u002Fstrong> で re-run。before \u002F after pair により \u003Cstrong>recipe 依存性\u003C\u002Fstrong>を数値で切り分け、卒論 §5.5 に measurement-driven な narrative を確立する。tuned 実測値は TBD (2026-07-22 執筆時点で training 稼働中、~2-3h)。","P1 G1 · Mip-NeRF 360 · bicycle · before \u002F after tuning","2026-07-22","stable","experiment","mixed",[21],1,[23,24,25,26,27,28,29,30],"mipnerf360","colmap","bicycle","real-data","recipe-tuning","before-after","pareto","post-hoc-eval","G1 Mip-NeRF 360 (7 月実行プラン、Pareto v2)",[33,34],"bicycle-mipnerf-30k","bicycle-mipnerf-tuned-3m",[36,37,38],"p1-axis1-phase-i-scene-adaptive","p1-msplat-baseline-spike","p1-mcmc-fix3-negative",{"summary":40,"rank":41,"verdict":42,"delta_psnr":43,"delta_wallclock":44},"\u003Cstrong>bicycle baseline (Phase I recipe 流用) = 15.549 dB \u002F Kerbl 2023 比 -9.7 dB\u003C\u002Fstrong>、tuned (refine.stop_iter=25000, capacity=3M、2 変数のみ isolated tuning) = \u003Cstrong>15.603 dB \u002F Δ +0.054 dB (実質 noise 圏)\u003C\u002Fstrong>。splats +55% (311k→483k) を追加しても PSNR は動かず、wallclock も ±0。\u003Cstrong>bottleneck は refine schedule \u002F capacity growth ではなく、より深い layer\u003C\u002Fstrong> (opacity reset cadence \u002F grad threshold \u002F unbounded outdoor scene handling)。\u003Cstrong>Case (b) scope-out 確定\u003C\u002Fstrong>: Apple Silicon native 実装の scope 境界は synthetic + object-centric に留まり、Mip-NeRF 360 実写は refine + capacity 調整では回収不能、future work として **本質的な refine schedule 再設計** or **msplat \u002F MCMC 系の densification 経路の再挑戦**が必要。\u003Cstrong>本 finding は §5.4 の MCMC 5 連敗 pattern と完全に一貫\u003C\u002Fstrong> (measurement で scope 境界を数値化、fix cost を明示化)。","high","case-b-scope-out-refine-capacity-not-bottleneck","baseline 15.549 → tuned 15.603 dB = **+0.054 dB (noise 圏)** \u002F Kerbl 2023 比 -9.65 dB (baseline とほぼ同 gap)","baseline 1h 49m → tuned 1h 42m = -7 min (±0、splats +55% を吸収)",[46,49,54,57,126,128,167,169,207,209,221,223,239,241],{"type":47,"text":48},"lead","G1 Mip-NeRF 360 実写 3-way Pareto の initial baseline (bicycle 30k) は post-hoc COLMAP eval で \u003Cstrong>PSNR 15.549 dB\u003C\u002Fstrong>、Kerbl 2023 paper 25.25 dB 比 \u003Cstrong>-9.7 dB\u003C\u002Fstrong> の大 gap。tuned recipe (refine.stop_iter=25000, capacity=3M、2 変数のみ isolated) で re-run した結果 \u003Cstrong>15.603 dB (Δ +0.054 dB = noise 圏)\u003C\u002Fstrong>、splats +55% を追加しても PSNR 動かず。\u003Cstrong>bottleneck は refine schedule \u002F capacity ではなく、より深い layer\u003C\u002Fstrong>と確定 (Case b scope-out)。§5.4 の MCMC 5 連敗と同 pattern の measurement-driven negative finding。",{"type":50,"label":51,"variant":52,"text":53},"callout","Headline (2 変数 isolated tuning で +0.054 dB のみ、Case b scope-out 確定)","warning","\u003Cstrong>bicycle baseline (Phase I recipe 流用)\u003C\u002Fstrong> = 15.549 dB (min 12.567 \u002F median 15.602 \u002F max 18.683 \u002F splats 311k \u002F 1h 49m)。\u003Cstrong>tuned (refine.stop_iter=25000, capacity=3M、他は同一)\u003C\u002Fstrong> = \u003Cstrong>15.603 dB (min 12.353 \u002F median 15.571 \u002F max 19.298 \u002F splats 483k \u002F 1h 42m)\u003C\u002Fstrong>。\u003Cstrong>Δ = +0.054 dB (実質 noise 圏、seed variance ±0.1 dB 内)\u003C\u002Fstrong>、splats +55% を追加しても PSNR 動かず。Kerbl 2023 比 gap は baseline -9.70 → tuned -9.65 とほぼ不変。\u003Cstrong>Case (b) scope-out 確定\u003C\u002Fstrong>: recipe の 2 変数 (refine.stop_iter \u002F capacity) は Mip-NeRF 360 実写での bottleneck ではない。より深い layer (opacity reset cadence \u002F grad threshold \u002F unbounded outdoor scene 特有の handling) が原因と推定される。§5.4 の MCMC 5 連敗 pattern と一貫、卒論 §5.5 の narrative は「Apple Silicon native 実装の scope 境界は synthetic + object-centric に留まる、実写 Mip-NeRF 360 は future work として refine schedule の本質的再設計または MCMC 系 densification の再挑戦が必要」で確定。",{"type":55,"text":56},"heading","1. Baseline 実測 (bicycle 30k, Phase I recipe 流用)",{"type":58,"columns":59,"align":63,"rows":66,"caption":125},"table",[60,61,62],"metric","value","備考",[64,65,64],"left","right",[67,71,75,79,83,87,91,95,99,103,107,111,114,117,121],[68,69,70],"config","`2026-07-22-bicycle-30k.toml`","Phase I recipe を実写に流用 (synthetic tuning のまま)",[72,73,74],"reflect.stop_iter","15,000","30k の 50%、synthetic の Phase H sweet spot 値を流用",[76,77,78],"trainer.capacity","1,500,000","M4 Max 36GB での保守的 cap",[80,81,82],"MCMC","OFF","G3 5 連敗確定済 (`p1-mcmc-fix3-negative`)",[84,85,86],"gt_convention","WhiteBg (legacy)","COLMAP は α=1.0 hardcode なので premultiplied と等価",[88,89,90],"train iterations","30,000","完走",[92,93,94],"train wallclock","1h 49m","M4 Max, 169 train views, 25 eval views",[96,97,98],"final splats","311,259","capacity 1.5M cap の 20.75% 使用",[100,101,102],"final loss (train)","3.03e-1","iter 30000 の training loss",[104,105,106],"**post-hoc PSNR (val 25 views)**","**15.549 dB**","**A-1.5 で追加した splat eval --format colmap の実測 mean**",[108,109,110],"  min PSNR","12.567 dB","",[112,113,110],"  median PSNR","15.602 dB",[115,116,110],"  max PSNR","18.683 dB",[118,119,120],"Kerbl 2023 paper 公称値","25.25 dB","Table 4, bicycle @ A6000 GPU, 標準 3DGS-CUDA",[122,123,124],"**Δ vs Kerbl 2023**","**-9.70 dB**","**大 gap、recipe 依存性を tuned で切り分ける根拠**","bicycle baseline (Phase I recipe 流用) の実測。post-hoc COLMAP eval は A-1.5 で追加した `splat eval --format colmap` サブコマンドを使用、`--colmap-downsample 4 --colmap-eval-stride 8` で訓練と同じ split を deterministic に再現。",{"type":55,"text":127},"2. Tuned recipe の設計 (before \u002F after 変数 2 個のみ弄る)",{"type":58,"columns":129,"align":134,"rows":135,"caption":166},[130,131,132,133],"parameter","baseline","tuned","変更理由",[64,65,65,64],[136,140,144,148,152,154,158,162],[137,73,138,139],"`refine.stop_iter`","**25,000**","実写 unbounded scene は synthetic より collapse が遅い、refine を最後 5k iter 前まで継続",[141,77,142,143],"`trainer.capacity`","**3,000,000**","Kerbl 2023 の bicycle 実装は 3M+ splats、baseline 311k は capacity 20% 使用で growth 制約なし",[145,146,146,147],"`refine.every`","100","据え置き、cadence 変数は isolate",[149,150,150,151],"`refine.opacity_reset_every`","3,000","据え置き",[80,81,81,153],"G3 5 連敗確定済で撤退",[155,156,156,157],"Adam lr","(same)","据え置き、synthetic と同一 recipe を維持",[159,160,161,151],"loss \u002F λ \u002F SSIM","l1_ssim \u002F 0.20 \u002F 7x1","同上",[163,164,164,165],"seed","42","同一 seed で variance を除去","before \u002F after で変更した変数は 2 個 (`refine.stop_iter` + `trainer.capacity`) のみ。他は全て同一。tuned が baseline を上回った場合、その +X dB は 2 変数の effect と直接 attribute できる (isolation)。",{"type":55,"text":168},"3. Tuned 実測 (2026-07-22 完走)",{"type":58,"columns":170,"align":171,"rows":172,"caption":206},[60,61,62],[64,65,64],[173,176,177,180,183,186,189,191,193,195,199,202],[68,174,175],"`2026-07-22-bicycle-tuned-3m.toml`","tuned recipe (2 変数のみ変更)",[88,89,90],[92,178,179],"1h 42m 17s","M4 Max, baseline 1h 49m から -7 min (splats +55% を吸収して同等)",[96,181,182],"483,019","capacity 3M cap の 16.1% 使用 (依然 growth 制約なし)、baseline 311k から +55%",[100,184,185],"2.68e-1","iter 30000 の training loss、baseline 3.03e-1 から微改善",[104,187,188],"**15.603 dB**","**A-1.5 の splat eval --format colmap の実測 mean**",[108,190,110],"12.353 dB",[112,192,110],"15.571 dB",[115,194,110],"19.298 dB",[196,197,198],"**Δ baseline → tuned**","**+0.054 dB**","**noise 圏 (seed variance ±0.1 dB 内)、2 変数調整は bottleneck ではない**",[122,200,201],"**-9.65 dB**","baseline -9.70 とほぼ同 gap",[203,204,205],"**splats efficiency**","**PSNR\u002Fsplat = 3.23e-5 dB\u002Fsplat (tuned) vs 5.00e-5 dB\u002Fsplat (baseline)**","**tuned は splat を +55% 追加したが PSNR\u002Fsplat 効率が -35%、over-densification の萌芽 (§5.4.3 opacity decay で観測した無駄 splat と同族)**","tuned run の実測結果。同 seed で baseline とは (`stop_iter`, `capacity`) の 2 変数のみ違い、その **Δ +0.054 dB (noise 圏)** が「refine schedule \u002F capacity は bottleneck ではない」ことの isolated evidence。追加した 172k splat (311k→483k) は PSNR に寄与せず、むしろ PSNR\u002Fsplat 効率を下げた (Phase D の over-densification 現象と同族)。",{"type":55,"text":208},"4. 解釈 (Case b scope-out 確定) と卒論 §5.5 への組み込み方針",{"type":210,"items":211},"list",[212,213,214,215,216,217,218,219,220],"\u003Cstrong>Case (b) 確定\u003C\u002Fstrong>: 2 変数 isolated tuning (refine.stop_iter 15k→25k, capacity 1.5M→3M) で Δ +0.054 dB (noise 圏) のみ。\u003Cstrong>bottleneck は refine schedule \u002F capacity growth ではない\u003C\u002Fstrong>ことが数値で切り分けられた。事前想定していた「refine 早期停止 + 低 capacity が synthetic tuning artifact」仮説は falsified。","\u003Cstrong>splats +55% で PSNR ±0\u003C\u002Fstrong>: baseline 311k → tuned 483k splats を追加しても PSNR は動かず、むしろ PSNR\u002Fsplat 効率が -35% に低下。これは Phase D (§5.4.3) で観測した \u003Cstrong>over-densification noise\u003C\u002Fstrong> と同族の現象。opacity decay を実写に適用すれば無駄 splat を trim できる可能性がある (future work 候補)。","\u003Cstrong>真の bottleneck 候補\u003C\u002Fstrong> (次に検証するなら primary suspect):","  (i) \u003Cstrong>opacity reset cadence\u003C\u002Fstrong>: baseline \u002F tuned とも 3000 iter ごと reset だが、実写の 169 view (synthetic の 100 view より多い) では reset 頻度が過剰かもしれない。\u003Cstrong>opacity_reset_every を 5000-7500 に延長する ablation\u003C\u002Fstrong> で ±0.5 dB 級の効果を期待。","  (ii) \u003Cstrong>grad threshold\u003C\u002Fstrong>: `refine.grad_threshold=2.0e-7` は synthetic 8 scene で tune された値。実写 unbounded outdoor は近景 \u002F 遠景の depth variance が桁違いに大きく、grad magnitude の分布も異なる。\u003Cstrong>scene-adaptive grad threshold or opacity-aware refine cadence\u003C\u002Fstrong> が必要。","  (iii) \u003Cstrong>unbounded outdoor scene 特有の handling\u003C\u002Fstrong>: bicycle scene は sky \u002F 遠景 mountain が dataset の半分を占めるが、SfM points 54k は近景 bicycle 中心に偏っている。\u003Cstrong>SfM init の scene coverage 改善\u003C\u002Fstrong> または \u003Cstrong>MCMC-style densify で init に依存しない growth\u003C\u002Fstrong> が必要。ただし MCMC は G3 5 連敗確定済 (`p1-mcmc-fix3-negative`)、msplat 再挑戦は 3-4 日工数。","\u003Cstrong>卒論 §5.5 の narrative\u003C\u002Fstrong>: 「Apple Silicon native 実装は Phase I recipe で synthetic 8 scene mean 33.56 dB \u002F brush +0.63 dB を維持する一方、Mip-NeRF 360 実写 (bicycle) では Kerbl 2023 25.25 dB に対し baseline -9.7 dB \u002F tuned -9.65 dB と gap をほぼ回収できず、本研究の scope 境界は \u003Cstrong>synthetic + object-centric\u003C\u002Fstrong> に留まる。2 変数 isolated tuning で recipe 変数ではないと数値で切り分けたことは、future work の primary suspect を opacity reset cadence \u002F grad threshold \u002F init coverage に絞る measurement-driven な貢献である」","\u003Cstrong>§5.4 との一貫性\u003C\u002Fstrong>: 本 finding は §5.4.4 (axis 1 ROI 上限実測)、§5.4.7 (Apple Silicon Metal 最適化 ROI 階層)、§5.4.X (MCMC 5 連敗) と同じ pattern を踏襲する — \u003Cstrong>事前 commit した仮説を実測で falsify し、次に必要な investigation の primary suspect を signals として encode する\u003C\u002Fstrong>。卒論全体の方法論的一貫性を守る。","\u003Cstrong>msplat \u002F brush との対比\u003C\u002Fstrong>: 本 finding は本実装単独の recipe 依存性を切り分けるもの。msplat \u002F brush の Mip-NeRF 360 実測は本セッションでは実施しない (msplat spike は synthetic のみ)。Pareto plot v2 は「本実装 baseline + tuned + Kerbl 2023 reference (paper 公称値)」の 3 系列で作る。他実装の Mip-NeRF 360 実測は Track D-1 (来週以降) の future work。",{"type":55,"text":222},"5. 実施ログ (2026-07-22)",{"type":210,"ordered":224,"items":225},true,[226,227,228,229,230,231,232,233,234,235,236,237,238],"07:06:37 — 4 scene sweep 開始 (bicycle → garden → counter → room)。runner: `bash splat\u002Fscripts\u002Frun-mipnerf360-4scene-30k.sh`。","08:55:48 — bicycle 30k 完走 (1h 49m)。garden 開始。","09:00 頃 — bicycle `result.toml.psnr = 0.0` を発見。train.rs の final eval は `transforms_val.json` 前提で COLMAP dataset では skip されると判明。","09:15 頃 — `splat eval` に `--format colmap` を追加 (A-1.5)。cargo build --release 成功、running training には影響なし。","09:20 頃 — bicycle final.ply で post-hoc COLMAP eval を実行 → **PSNR 15.549 dB** 確定 (Kerbl 2023 比 -9.7 dB)。","09:30 頃 — 「残り 3 scene 同 recipe で回すのは ROI 低」と判断、sweep 中断。bicycle 1 scene に絞って tuned re-run に方針転換 (案 B)。","09:40 頃 — tuned config `2026-07-22-bicycle-tuned-3m.toml` を作成、`splat check` で validate 通過、`Bash(run_in_background)` で launch (PID 41409、log: `runs-logs\u002Fbicycle-tuned-3m-20260722-094055.log`)。","09:40 頃 — tuned config launch (1 回目、Bash run_in_background)。iter 16500 で unintentional kill。","10:37 頃 — tuned run 再起動 (2 回目、同 config、Bash run_in_background)。iter 10000 で再び unintentional kill。","11:05 頃 — nohup + disown で detached 起動 (3 回目、PPID=1)、harness kill 対象外。以降 iter 15000 \u002F 20000 \u002F 25000 全て通過。","12:48 頃 — tuned run 完走 (1h 42m 17s、final splats 483k、final loss 0.268)。","12:50 頃 — post-hoc COLMAP eval 実行、**PSNR mean 15.603 dB \u002F Δ baseline +0.054 dB (noise 圏)** 確定 → **Case (b) scope-out**。","12:55 頃 — 本 finding doc + §5.5 LaTeX を Case (b) narrative で更新。Pareto plot v2 (bicycle before\u002Fafter) 生成。",{"type":55,"text":240},"6. 関連",{"type":210,"items":242},[243,244,245,246,247,248,249,250,251,252,253,254,255],"baseline config: \u003Ccode>splat\u002Fconfigs\u002F2026-07-22-bicycle-30k.toml\u003C\u002Fcode>","tuned config: \u003Ccode>splat\u002Fconfigs\u002F2026-07-22-bicycle-tuned-3m.toml\u003C\u002Fcode>","baseline result: \u003Ccode>splat\u002Fruns\u002Fbicycle-mipnerf-30k\u002Fresult.toml\u003C\u002Fcode> (psnr=0.0、post-hoc で 15.549 dB) + \u003Ccode>final.ply\u003C\u002Fcode> (73 MB, 311k splats)","baseline log: \u003Ccode>runs-logs\u002Fbicycle-mipnerf-30k-20260722-070637.log\u003C\u002Fcode>","tuned log: \u003Ccode>runs-logs\u002Fbicycle-tuned-3m-20260722-094055.log\u003C\u002Fcode>","post-hoc eval 実装: \u003Ccode>splat\u002Fcrates\u002Fsplat-cli\u002Fsrc\u002Fcmd\u002Feval.rs\u003C\u002Fcode> + \u003Ccode>src\u002Fmain.rs\u003C\u002Fcode> (A-1.5)","harvest スクリプト (post-hoc eval 統合済): \u003Ccode>splat\u002Fscripts\u002Fharvest-mipnerf-4scene.sh\u003C\u002Fcode>","Pareto v2 plot 生成器: \u003Ccode>docs\u002Fstrategy\u002Fmake_pareto_v2_12scene.py\u003C\u002Fcode> (現状は 4 scene 用、bicycle 2 点対応に修正予定)","Phase I (synthetic 対比): \u003Ccode>docs\u002Ffindings\u002Fp1-axis1-phase-i-scene-adaptive.toml\u003C\u002Fcode>","msplat spike (synthetic 対比): \u003Ccode>docs\u002Ffindings\u002Fp1-msplat-baseline-spike.toml\u003C\u002Fcode>","G3 5 連敗 (同 pattern の negative finding): \u003Ccode>docs\u002Ffindings\u002Fp1-mcmc-fix3-negative.toml\u003C\u002Fcode>","7 月実行プラン: \u003Ccode>docs\u002Fstrategy\u002F2026-07-execution-plan.md\u003C\u002Fcode>","残タスク list: \u003Ccode>docs\u002Fstrategy\u002F2026-07-final-tasklist.md\u003C\u002Fcode>",[257,267],{"id":33,"title":33,"subtitle":258,"date":16,"workspace":259,"tags":260,"verdict":262,"psnr":263,"psnr_unit":-1,"wallclock":264,"splats":265,"summary_url":266,"detail_path":266},"G1 Mip-NeRF 360 実写 3-way Pareto の 1 point — bicycle 30k full run","splat",[23,24,25,261,29],"30k","partial",0,"1h 49m 9s",311259,"\u002Fruns\u002Fbicycle-mipnerf-30k\u002F",{"id":34,"title":34,"subtitle":268,"date":16,"workspace":259,"tags":269,"verdict":262,"psnr":263,"psnr_unit":-1,"wallclock":270,"splats":271,"summary_url":272,"detail_path":272},"G1 Mip-NeRF 360 実写 Pareto — bicycle 30k tuned recipe (refine.stop_iter=25000, capacity=3M)",[23,24,25,261,29,132],"1h 42m 16s",483019,"\u002Fruns\u002Fbicycle-mipnerf-tuned-3m\u002F",[274,306,326],{"id":38,"title":275,"date":276,"status":17,"polarity":277,"category":278,"axes":279,"tags":280,"task_code":288,"related_runs":289,"delta_psnr":301,"delta_wallclock":302,"rank":41,"verdict":303,"impact_summary":304,"detail_path":305},"G3 MCMC fix3 — 5 連敗確定、scale_l1 犯人 + relocation 30k で 14.6 dB 頭打ち","2026-07-21","negative","negative-finding",[21],[281,282,283,277,284,285,286,287],"mcmc","g3","fix3","5-consecutive-failure","scale-eigen-l1","ablation","lego-30k","G3 MCMC fix3 (7 月実行プラン)",[290,291,292,293,294,295,296,297,298,299,300],"lego-mcmc-fix3-noise5e1-smoke","lego-mcmc-fix3-noise5e2-smoke","lego-mcmc-fix3-noise5e3-smoke","lego-mcmc-fix3-noise5e1-iter2000","lego-mcmc-fix3-noise5e2-iter2000","lego-mcmc-fix3-noise5e3-iter2000","lego-mcmc-fix3-ablation-A-baseline-iter2000","lego-mcmc-fix3-ablation-B-scalel1only-iter2000","lego-mcmc-fix3-ablation-C-relocationonly-iter2000","lego-mcmc-fix3-final-relocation-only-30k","lego-mcmc-fix3-final-noise-relocation-30k","Phase I 36.11 dB 比 -21 dB (relocation-only) \u002F -21 dB (noise-relocation)、fix2 17.38 dB 比 -2.7 dB (更に劣化)","Phase I 11m48s 比 +11x (relocation-only 2h12m) \u002F +13x (noise-relocation 2h38m)","5-consecutive-failure-scope-out","\u003Cstrong>G3 MCMC fix3 = 5 連敗確定\u003C\u002Fstrong>。fix2 (17.38 dB) 敗因の calibration ズレを、既実装の 3 欠陥修正版で再挑戦。\u003Cstrong>500 iter smoke で crash なし完走\u003C\u002Fstrong>、\u003Cstrong>2000 iter で 10-12 dB 頭打ち\u003C\u002Fstrong>、\u003Cstrong>ablation で scale_eigen_l1 の毒性を -4.4 dB と測定\u003C\u002Fstrong>、\u003Cstrong>scale_l1 除去した relocation + growth schedule の 30k full 2 パターンで PSNR 14.65 \u002F 14.69 dB\u003C\u002Fstrong> (Phase I baseline 36.11 dB 比 -21 dB)。犯人は複合: (1) scale_eigen_l1 実装が過剰罰化 or 符号ミス、(2) relocation の Eq.9 probability-preserving 実装が本実装の SoA parameterization と結合していない (splats 1.3-1.8M まで肥大化するが PSNR が付いてこない)。\u003Cstrong>7 月実行プラン G3 の timebox 判定基準通り即撤退\u003C\u002Fstrong>、卒論 §5.4 negative finding として 5 連敗目に確定、future work 送り。\u003Cstrong>ただし本 negative finding は従来の 4 連敗より格段に強い evidence\u003C\u002Fstrong> (component ablation で犯人を数値で切り分け、30k full の decisive failure まで到達)、卒論 measurement contribution として §5.4 で厚く記述可能。","\u002Ffindings\u002Fp1-mcmc-fix3-negative\u002F",{"id":37,"title":307,"date":308,"status":17,"polarity":19,"category":18,"axes":309,"tags":312,"task_code":319,"related_runs":320,"delta_psnr":321,"delta_wallclock":322,"rank":41,"verdict":323,"impact_summary":324,"detail_path":325},"msplat v1.1.3 baseline spike — M4 Max 実測、Lego 30K で 3m40s \u002F 35.48 dB、materials で densify 破綻","2026-06-26",[21,310,311],2,3,[313,131,314,315,316,317,318],"msplat","apple-silicon","m4-max","nerf-synthetic","direct-comparison","thesis-ready","P1 msplat spike (critical path)",[],"Lego: 本実装 +0.78 dB (Phase H 12500 vs msplat 30K)、materials: 本実装 +21.17 dB (Phase I 12500 vs msplat 7K broken)","Lego: msplat -3.2× (msplat 3m40s vs Phase H 11m48s)、chair 7K のみ msplat -63× (msplat 17.58s vs Phase I 18m35s)","critical-path-complete-A-confirmed","\u003Cstrong>msplat (rayanht) v1.1.3 を M4 Max で実測完了\u003C\u002Fstrong>。Lego 30K で 219.8s (3m40s) \u002F 35.48 dB \u002F 208k splats、chair 7K で 17.58s \u002F 33.81 dB を確認。\u003Cstrong>本実装 Phase H lego (11m48s \u002F 36.26 dB) に対し msplat は wallclock 3.2× 高速、ただし PSNR は -0.78 dB\u003C\u002Fstrong>。materials 7K では PSNR 8.80 \u002F splats 367 と densification が壊れた — msplat の scene robustness 弱点が判明し、本実装の \u003Cstrong>universal な density 動作\u003C\u002Fstrong>が明確な差別化軸として確定。\u003Cstrong>本実装 Phase I 8 scene のうち 5 scene が msplat lego 30K の 35.48 dB を上回る\u003C\u002Fstrong> (Lego 36.26 \u002F chair 35.88 \u002F hotdog 37.37 \u002F mic 36.62 \u002F drums は 27 のみ)。Pareto plot で msplat は左上 (速い + 中品質)、本実装は右側 (遅い + 高品質 + 全 scene robust) と vector が直交。本 spike により \u003Cstrong>A (Reframing narrative) の killer figure が確定\u003C\u002Fstrong>、\u003Cstrong>B (G.2 backport) は引き続き NO-GO\u003C\u002Fstrong> (msplat が既に kernel fusion + pre-allocated bins まで実装済、本実装が追いつく工数は大)、\u003Cstrong>E (msplat 著者接触) の materials bug 報告は良い PR ネタ\u003C\u002Fstrong>。","\u002Ffindings\u002Fp1-msplat-baseline-spike\u002F",{"id":36,"title":327,"date":328,"status":17,"polarity":329,"category":330,"axes":331,"tags":332,"task_code":341,"related_runs":342,"delta_psnr":345,"delta_wallclock":346,"rank":41,"verdict":347,"impact_summary":348,"detail_path":349},"Phase I scene-adaptive iter budget — **STRONG Pareto improvement** 確定 (8 scene mean +0.077 dB at -24% wallclock vs Phase D)","2026-05-26","positive","design",[21],[333,334,335,336,337,338,339,340],"p1-axis1","phase-i","scene-adaptive","pareto-front","stop-iter","sh-progressive","universal-improvement","calibration","P1 Phase I",[343,344],"chair-phase-i-adaptive-12500","materials-phase-i-adaptive-12500","**+0.077 dB** (8 scene mean vs Phase D 30k)、+0.701 dB (vs brush)、-0.031 dB (vs G.3 30k = within noise)","**-24.4%** (vs Phase D)、**-31.6%** (vs G.3 30k)、both Pareto dimensions improved","scene-adaptive-pareto-confirmed","Phase H で Lego stop_iter=12500 が Pareto sweet spot だが ficus\u002Fmic で大幅 fail (-4.12 \u002F -6.03 dB) と scene-dependent 確定後、**Phase I で chair \u002F materials @ stop_iter=12500 も fast converger と確認** (G.3 30k 比 -0.05〜-0.07 dB のみで essentially same)。これにより \u003Cstrong>scene-adaptive iter budget\u003C\u002Fstrong> (Lego\u002Fchair\u002Fmaterials @ 12500、ficus\u002Fdrums\u002Fhotdog\u002Fmic\u002Fship @ 30000) を構成、\u003Cstrong>既存 8 scene 全データ点が揃った状態で 8 scene mean を算出\u003C\u002Fstrong>。\u003Cstrong>結果\u003C\u002Fstrong>: scene-adaptive 8 scene mean **33.561 dB** \u002F total wallclock **3h 50m 10s**、Phase D 30k 比 \u003Cstrong>+0.077 dB \u002F -24.4% wallclock\u003C\u002Fstrong>、G.3 30k 比 \u003Cstrong>-0.031 dB (within noise) \u002F -31.6% wallclock\u003C\u002Fstrong>、brush 比 \u003Cstrong>+0.701 dB\u003C\u002Fstrong>。\u003Cstrong>Pareto front の両軸で improvement\u003C\u002Fstrong>: Phase D を quality + speed 両方で dominate、G.3 30k quality を 1\u002F3 短時間で達成。**axis 1 contribution の最終形**: kernel-level 直叩きではなく \u003Cstrong>scene-adaptive iter budget + sh_progressive + opacity_decay の組み合わせ\u003C\u002Fstrong>が Apple Silicon native Metal 最適化の universal Pareto improvement。卒論 §5.4.7 末尾 + §6 future work で本 Phase I を明示、scene-adaptive を新 universal default として推奨。","\u002Ffindings\u002Fp1-axis1-phase-i-scene-adaptive\u002F",1790227326536]