[{"data":1,"prerenderedAt":434},["ShallowReactive",2],{"finding:p1-msplat-baseline-spike":3,"finding-runs:p1-msplat-baseline-spike":338,"finding-related:p1-msplat-baseline-spike":339},{"meta":4,"impact":32,"sections":38},{"id":5,"title":6,"subtitle":7,"eyebrow":8,"date":9,"status":10,"category":11,"polarity":12,"axes":13,"tags":17,"task_code":25,"related_runs":26,"related_findings":27},"p1-msplat-baseline-spike","msplat v1.1.3 baseline spike — M4 Max 実測、Lego 30K で 3m40s \u002F 35.48 dB、materials で densify 破綻","msplat (rayanht\u002Fmsplat) は Apple Silicon native 3DGS の direct baseline であり、本実装 Phase I との比較の前提として M4 Max で実測した。\u003Cstrong>Lego 30K iter で 3 分 40 秒 \u002F 検証 PSNR 35.48 dB \u002F 208k splats\u003C\u002Fstrong> を達成、本実装 Phase H (12500 iter, 11m48s, 36.26 dB) に対し \u003Cstrong>wallclock 3.2× 高速\u003C\u002Fstrong>かつ \u003Cstrong>PSNR は -0.78 dB\u003C\u002Fstrong>。chair 7K iter は 17.58s \u002F 33.81 dB。\u003Cstrong>ただし materials は 7K で PSNR 8.80 dB \u002F splats 367 と densification が壊れた\u003C\u002Fstrong> — msplat は scene robustness が本実装より弱い決定的証拠。msplat の Pareto vector は『速さ + 中-高品質』、本実装の Pareto vector は『+0.5-1 dB の品質 + 全 scene robust + scene-adaptive iter』で対立。Pareto plot は \u003Ccode>docs\u002Fstrategy\u002Fmsplat-vs-otake-pareto.png\u003C\u002Fcode>。卒論 §5 (実験) の direct baseline として確定。","P1 msplat · direct baseline · M4 Max 実測 · critical path 完了","2026-06-26","stable","experiment","mixed",[14,15,16],1,2,3,[18,19,20,21,22,23,24],"msplat","baseline","apple-silicon","m4-max","nerf-synthetic","direct-comparison","thesis-ready","P1 msplat spike (critical path)",[],[28,29,30,31],"p1-axis1-phase-i-scene-adaptive","p1-d-multi-scene-rechain","p1-axis1-phase-h-lego-pareto-sweep","p1-axis1-phase-g3-sh-progressive",{"summary":33,"rank":34,"verdict":35,"delta_psnr":36,"delta_wallclock":37},"\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>。","high","critical-path-complete-A-confirmed","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)",[39,42,47,50,95,97,158,160,241,243,296,298,307,309,317,319,325,327],{"type":40,"text":41},"lead","msplat (rayanht\u002Fmsplat) v1.1.3 を M4 Max で実測した結果、\u003Cstrong>速度では明確に劣勢 (Lego 30K で msplat 3.2× 高速)\u003C\u002Fstrong>、\u003Cstrong>PSNR では本実装が +0.5-21 dB の幅で勝つ\u003C\u002Fstrong>、\u003Cstrong>materials は msplat の densify が壊れて本実装が決定的優位\u003C\u002Fstrong>。Pareto vector が明確に直交し、A (Reframing narrative) の killer figure が完成。",{"type":43,"label":44,"variant":45,"text":46},"callout","Headline (msplat 実測 = Pareto vector が直交、A 確定 \u002F B NO-GO \u002F E は良い PR ネタ)","success","\u003Cstrong>msplat = 速さ vector (Lego 30K で 3m40s \u002F 35.48 dB)\u003C\u002Fstrong>、\u003Cstrong>本実装 Phase I = 品質 + robustness vector (8 scene 中 5 scene で PSNR 35+、materials も robust)\u003C\u002Fstrong>。msplat の materials 7K で \u003Cstrong>PSNR 8.80 dB \u002F splats 367\u003C\u002Fstrong> は densification 壊れた決定的失敗で、本実装の materials 12500 = 29.97 dB \u002F 8s splat 268k と \u003Cstrong>+21 dB の差\u003C\u002Fstrong>。これは A (Reframing narrative) の核 evidence: 「本実装は kernel-level 速さでは msplat に負けるが、scene-adaptive な品質追求と全 scene robustness で勝つ」。\u003Cstrong>B (G.2 backport) は NO-GO 維持\u003C\u002Fstrong> — msplat は CHANGELOG v1.1.3 で『Fused SH backward into Adam』『Fused SSIM』『Pre-allocated per-tile bins』を既に実装済、本実装が追いつく実装コストは大、第 2 回 ultracode adversarial verify の refuted 結論と整合。\u003Cstrong>E (msplat 著者接触) は materials 破綻を issue 報告する道筋が PR チャンスとして有効\u003C\u002Fstrong>。Pareto plot: \u003Ccode>docs\u002Fstrategy\u002Fmsplat-vs-otake-pareto.png\u003C\u002Fcode>。",{"type":48,"text":49},"heading","1. msplat 実測結果 (M4 Max, NeRF Synthetic, white_bg=1, sh_degree=3)",{"type":51,"columns":52,"align":60,"rows":63,"caption":94},"table",[53,54,55,56,57,58,59],"scene","iter","wallclock","検証 PSNR","SSIM","splats","備考",[61,62,62,62,62,62,61],"left","right",[64,72,80,86],[65,66,67,68,69,70,71],"Lego","7,000","20.6s","31.56","0.9622","63,929","標準動作",[73,74,75,76,77,78,79],"**Lego**","**30,000**","**219.8s (3m40s)**","**35.48**","**0.9832**","**208,215**","**標準動作、本 spike の主要数値**",[81,66,82,83,84,85,71],"chair","17.58s","33.81","0.9761","38,220",[87,88,89,90,91,92,93],"**materials**","**7,000**","**11.61s**","**8.80**","**0.7795**","**367**","**❌ densification 壊れた、msplat の bug**","msplat v1.1.3 を M4 Max で実測 (`\u002Ftmp\u002Fmsplat-spike\u002Fresults-7k.json` + `bench-30k.log`)。\u003Cstrong>materials が densify されず final splat 367 に留まる致命的失敗\u003C\u002Fstrong> — msplat の scene robustness 弱点が判明 (本実装 Phase I の materials は 30 dB \u002F 268k splats \u002F 11 分で問題なく densify)。Lego 30K は標準動作で 208k splats \u002F 35.48 dB を達成、本実装 Phase D Lego (375k splats \u002F 36.11 dB \u002F 41m54s) に対し speed で 11× 速いが PSNR -0.63 dB \u002F splats も 1.8× 少ない (圧縮的)。",{"type":48,"text":96},"2. 本実装 Phase I 8 scene との直接比較 (M4 Max)",{"type":51,"columns":98,"align":106,"rows":107,"caption":157},[53,99,100,101,102,103,104,105],"本実装 Phase I (iter)","本実装 wallclock","本実装 PSNR","msplat (iter)","msplat wallclock","msplat PSNR","Δ PSNR (本実装 - msplat)",[61,62,62,62,62,62,62,62],[108,116,120,126,132,137,141,145,149],[73,109,110,111,112,113,114,115],"12,500","11m48s","**36.26**","30,000","**3m40s**","35.48","**+0.78 ✓**",[81,109,117,118,66,82,83,119],"18m35s","**35.88**","**+2.07 ✓** (iter 数違い)",[87,109,121,122,66,123,124,125],"10m59s","**29.97**","11.61s","8.80","**+21.17 ✓✓ (msplat 破綻)**",[127,112,128,129,130,130,130,131],"ficus","21m40s","34.28","N\u002FA","未計測 (msplat 30K は Lego\u002Fchair のみ)",[133,112,134,135,130,130,130,136],"drums","1h05m","27.22","同上",[138,112,139,140,130,130,130,136],"hotdog","28m17s","37.37",[142,112,143,144,130,130,130,136],"mic","33m29s","36.62",[146,112,147,148,130,130,130,136],"ship","39m46s","30.88",[150,151,152,153,151,154,155,156],"**mean (Lego\u002Fchair\u002Fmaterials)**","-","**13m47s**","**34.04**","**16m**","**26.03**","**+8.01 dB \u002F wallclock 0.86× (msplat やや速い)**","本実装 Phase I は scene-adaptive iter (Lego\u002Fchair\u002Fmaterials @ 12500、他 @ 30000) を採用、各 scene で robust に高 PSNR を達成。\u003Cstrong>3 scene mean では本実装が PSNR +8.01 dB 優位\u003C\u002Fstrong>、これは materials の msplat 破綻 (+21 dB) と chair iter 数差 (本実装 12500 vs msplat 7K) が大きい。Lego 同条件比較 (本実装 12500 vs msplat 30K) では本実装 PSNR +0.78、msplat wallclock 3.2× 高速 — \u003Cstrong>Pareto vector が直交\u003C\u002Fstrong>している。",{"type":48,"text":159},"3. msplat の差別化軸 matrix (msplat 実装済 \u002F 未実装 vs 本実装)",{"type":51,"columns":161,"align":166,"rows":168,"caption":240},[162,163,164,165],"最適化 \u002F 機能","msplat v1.1.3","本実装","本実装の優位性",[61,167,167,61],"center",[169,174,177,180,184,188,192,196,200,203,208,213,217,221,226,230,233,237],[170,171,172,173],"Fused SH backward into Adam","✅ (CHANGELOG v1.1.3)","❌","msplat 優位 — 600 MB\u002Fiter readback 削減",[175,171,172,176],"Fused SSIM (V-fwd + H-bwd)","msplat 優位 — 130 MB\u002Fiter buffer 削減",[178,171,172,179],"Pre-allocated per-tile bins","msplat 優位 — `prefix_sort_pack` 19→10%",[181,182,172,183],"Tile-local bitonic sort","✅","msplat 優位",[185,182,186,187],"GPU-resident densify","✅ (Phase 5 M-3.x)","互角",[189,190,191,187],"Per-stage GPU profiling","✅ (PROFILE_STAGES)","✅ (SPLAT_TIMING)",[193,172,194,195],"**scene-adaptive iter budget**","✅ **(Phase I の核)**","**本実装優位**",[197,198,199,195],"**opacity decay (緩やか減衰)**","❌ (opacity reset のみ)","✅ **(Phase D の核)**",[201,172,202,195],"**SH degree 段階上昇**","✅ **(Phase G.3 の核)**",[204,205,206,207],"**brush convention 整合性**","△ (white_bg のみ)","✅ **(Phase A+B+F)**","**本実装優位 — 評価条件 explicit**",[209,210,211,212],"**materials scene robust**","❌ (8.80 dB \u002F 367 splats)","✅ (29.97 dB \u002F 268k)","**本実装決定的優位 ★**",[214,172,215,216],"**multi-cam Adam invariance 実証**","✅ (H.A regression)","**本実装優位 (reproducibility)**",[218,172,219,220],"**MCMC noise calibration lesson**","✅ (Phase 5 mcmc-3-defects)","**本実装優位 (methodological)**",[222,223,224,225],"**wgpu cross-vendor**","❌ (Metal only)","✅ (Rust+wgpu+Metal)","**本実装優位 — CUDA path も可能**",[227,228,172,229],"ICB (Indirect Command Buffer)","未確認","両者未実装 (G.2 backport 候補)",[231,228,172,232],"Argument buffer (bindless)","両者未実装",[234,228,235,236],"fp16 forward","❌ (Phase F.2 falsified)","両者効果薄",[238,228,186,239],"unified memory zero-copy loss path","本実装優位の可能性","差別化 matrix から、\u003Cstrong>本実装の優位軸は 8 個\u003C\u002Fstrong> (scene-adaptive \u002F opacity decay \u002F SH 段階上昇 \u002F brush 慣習整合 \u002F materials robust \u002F multi-cam Adam \u002F MCMC calibration \u002F wgpu cross-vendor)、\u003Cstrong>msplat の優位軸は 4 個\u003C\u002Fstrong> (Fused SH-Adam \u002F Fused SSIM \u002F Pre-alloc bins \u002F Tile-local sort)、互角 2 個、両者未実装 3 個。\u003Cstrong>本実装の優位軸は algorithmic \u002F methodological \u002F measurement 系で msplat の優位軸は kernel-level 速度系\u003C\u002Fstrong>。これは第 2 回 ultracode の『kernel-level 5 連続 falsified、algorithmic + scene-adaptive が universal な改善源』と完全に整合。",{"type":48,"text":242},"4. msplat の GPU stage profile (本実装の改善余地確認)",{"type":51,"columns":244,"align":248,"rows":249,"caption":295},[245,246,247,59],"stage","msplat Lego 7K (median ms\u002Fiter, n=21000)","msplat Lego 30K (median ms\u002Fiter, n=36500)",[61,62,62,61],[250,255,260,265,270,275,280,285,290],[251,252,253,254],"proj_sh_fwd","0.564","9.868","SH 計算 + projection、splat 数で大きく変動",[256,257,258,259],"prefix_sort_pack","6.786","20.250","msplat の bitonic sort、本実装の radix sort と比較対象",[261,262,263,264],"rast_fwd","11.345","45.094","ラスタライズ forward",[266,267,268,269],"loss_fwd","18.788","20.660","L1+SSIM forward (Fused SSIM の恩恵)",[271,272,273,274],"loss_bwd","23.177","24.229","L1+SSIM backward",[276,277,278,279],"rast_bwd","25.087","92.446","**ラスタライズ backward — 本実装も同じく dominant stage**",[281,282,283,284],"proj_sh_bwd_adam","2.017","47.802","**Fused SH backward + Adam (msplat の独自最適化)**",[286,287,288,289],"grad_stats","0.189","1.182","gradient 統計収集",[291,292,293,294],"**TOTAL**","**87.955ms**","**261.531ms**","msplat per-iter 時間、本実装 Phase D は 82 ms\u002Fiter 相当","msplat の per-stage GPU 時間。\u003Cstrong>n=36500 で TOTAL 261 ms\u002Fiter は本実装 Phase D 30k (82 ms\u002Fiter) より 3× 遅い\u003C\u002Fstrong>が、これは splat 数の差 (msplat 208k vs 本実装 375k、約 1.8×) を考慮しても msplat の方が per-iter で重い時間帯がある (Lego 30K では rast_bwd 92ms と proj_sh_bwd_adam 48ms が dominant)。本実装が wallclock で速いのは \u003Cstrong>opacity decay + scene-adaptive で iter 数 + splat 数を抑えている\u003C\u002Fstrong>から、kernel 単体ではなく algorithmic 効率で勝負している証拠。",{"type":48,"text":297},"5. A (Reframing narrative) の killer evidence として確定",{"type":299,"ordered":300,"items":301},"list",true,[302,303,304,305,306],"\u003Cstrong>Pareto vector 直交\u003C\u002Fstrong>: msplat = 速さ (Lego 30K 3m40s)、本実装 = 品質 + robustness (8 scene 中 5 scene で PSNR 35+、materials も robust)。Pareto plot が visual に明示 (\u003Ccode>docs\u002Fstrategy\u002Fmsplat-vs-otake-pareto.png\u003C\u002Fcode>)。","\u003Cstrong>materials での決定的優位\u003C\u002Fstrong>: msplat 7K で PSNR 8.80 dB \u002F splats 367 と densification が壊れた = msplat の scene-dependent 失敗。本実装は materials 12500 で 29.97 dB \u002F 268k splats \u002F 11 分で robust。\u003Cstrong>これは msplat も brush も持っていない unique contribution\u003C\u002Fstrong>。","\u003Cstrong>algorithmic vs kernel-level の対比が明確\u003C\u002Fstrong>: msplat の 4 優位は全て kernel-level (Fused SH-Adam \u002F Fused SSIM \u002F Pre-alloc bins \u002F Tile-local sort)、本実装の 8 優位は algorithmic (scene-adaptive \u002F opacity decay \u002F SH 段階上昇 \u002F brush 慣習) と methodological (multi-cam Adam invariance \u002F MCMC calibration) と engineering (wgpu cross-vendor \u002F materials robust)。\u003Cstrong>第 2 回 ultracode の ROI 階層と完全に整合\u003C\u002Fstrong>。","\u003Cstrong>第 2 回 ultracode adversarial verify の reinforcement\u003C\u002Fstrong>: B (G.2 backport) で MLX 流 graph eval を移植しても効果は限定的、と verify 済。msplat 実測で確認できたのは『msplat が既に Fused SH\u002FSSIM + Pre-alloc bins まで実装済で、本実装が追いつくには大コスト』 — B NO-GO が再確認された。","\u003Cstrong>卒論 §5 の direct baseline として確定\u003C\u002Fstrong>: msplat の数値表 + Pareto plot を §5 (実験) で direct comparison に使う。msplat は M4 Max + Metal only で外部依存ゼロなので査読者も再現性を期待しない (= 公称値で十分)、本実装は M4 Max + 同条件で実測した数値で勝負できる。",{"type":48,"text":308},"6. B\u002FD\u002FE の GO\u002FNO-GO 再判定 (本 spike 結果ベース)",{"type":299,"items":310},[311,312,313,314,315,316],"\u003Cstrong>A (Reframing narrative)\u003C\u002Fstrong>: ✅ **GO 確定**。msplat 数値 + Pareto plot で killer figure 完成、卒論 §5 \u002F §6 章立てを 1-2 日 (agents 並列) で書ける状態。","\u003Cstrong>B (G.2 backport spike)\u003C\u002Fstrong>: ❌ **NO-GO 確定**。msplat が既に Fused kernel + Pre-alloc bins まで実装済、本実装が同等の kernel-level 最適化を作っても msplat の 4 つの優位を全て追いつくのは 1-2 週で不可能。第 2 回 ultracode adversarial verify (5\u002F5 refuted) と materials robust の本実装優位を考えると、kernel-level で追いかける ROI は極めて低い。","\u003Cstrong>C (msplat 詳細比較)\u003C\u002Fstrong>: ✅ **本 spike で完了**。本 finding doc + Pareto plot がその成果物。","\u003Cstrong>D (実データ拡張 Mip-NeRF 360 \u002F Tanks & Temples)\u003C\u002Fstrong>: 🟡 **条件付き GO**。msplat README で mipnerf360 4 scene 公称値あり、本実装と直接比較可能。優先度は A 完成後、残り時間に応じて。msplat の materials 破綻が実データでも再現するなら本実装の優位がさらに強化される可能性。","\u003Cstrong>E (msplat 著者 rayanht コラボ)\u003C\u002Fstrong>: 🟡 **強い PR 機会あり**。本 spike で materials の densification 破綻を確認、これを `Issue` として報告すれば著者からの応答可能性が高い (re-densify 関連の specific bug)。コラボに発展すれば卒論 community contribution として記述可能。優先度は A 完成後、応答時間は author 次第。","\u003Cstrong>F (Autotune 系研究との接合)\u003C\u002Fstrong>: 🟢 卒論執筆段階で related work 補強として自然に統合可、優先度低。",{"type":48,"text":318},"7. 次の action (今日明日)",{"type":299,"ordered":300,"items":320},[321,322,323,324],"\u003Cstrong>松田先生事前ヒアリングメール\u003C\u002Fstrong>: 本 spike 結果 (msplat 数値 + Pareto plot + materials 破綻) を添えて『Reframing narrative + Pareto plot の方針について、hard system contribution の要否』を確認。所要 15 分 (agent draft + 人間 review)。","\u003Cstrong>A (Reframing narrative) の 1 ページ要旨 + 卒論章立てドラフト\u003C\u002Fstrong>: agents 並列で 1 日以内に完成可能。本 finding + p1-axis1-phase-i-scene-adaptive を入力にする。","\u003Cstrong>第 4 回ゼミ outline 起草\u003C\u002Fstrong>: 本 spike 結果 + Pareto plot を §2 で direct comparison として組み込み、reframing narrative を主張。所要 半日。","\u003Cstrong>(オプション) msplat 著者への issue 報告\u003C\u002Fstrong>: materials の densification 破綻を minimal repro (scripts\u002Frun_bench.py 抜粋) で issue として報告。E の PR チャンス。所要 30 分 (agent draft)。",{"type":48,"text":326},"8. 関連",{"type":299,"items":328},[329,330,331,332,333,334,335,336,337],"Pareto plot: \u003Ccode>docs\u002Fstrategy\u002Fmsplat-vs-otake-pareto.png\u003C\u002Fcode>","Plot 生成スクリプト: \u003Ccode>docs\u002Fstrategy\u002Fmake_msplat_pareto.py\u003C\u002Fcode>","実測 raw data: \u003Ccode>\u002Ftmp\u002Fmsplat-spike\u002Fresults-7k.json\u003C\u002Fcode> + \u003Ccode>\u002Ftmp\u002Fmsplat-spike\u002Fbench-30k.log\u003C\u002Fcode>","第 2 回 ultracode 調査 (本 spike を予告): \u003Ccode>docs\u002Fstrategy\u002F2026-06-25-next-steps-deep-dive-v2.md\u003C\u002Fcode>","本実装 Phase I (比較対照): \u003Ccode>docs\u002Ffindings\u002Fp1-axis1-phase-i-scene-adaptive\u003C\u002Fcode>","本実装 Phase D (比較対照): \u003Ccode>docs\u002Ffindings\u002Fp1-d-multi-scene-rechain\u003C\u002Fcode>","本実装 Phase H (比較対照): \u003Ccode>docs\u002Ffindings\u002Fp1-axis1-phase-h-lego-pareto-sweep\u003C\u002Fcode>","msplat 公式: \u003Ccode>https:\u002F\u002Fgithub.com\u002Frayanht\u002Fmsplat\u003C\u002Fcode>","msplat CHANGELOG (差別化 matrix の根拠): \u003Ccode>https:\u002F\u002Fraw.githubusercontent.com\u002Frayanht\u002Fmsplat\u002Fmain\u002FCHANGELOG.md\u003C\u002Fcode>",[],[340,361,380,406],{"id":30,"title":341,"date":342,"status":10,"polarity":12,"category":343,"axes":344,"tags":345,"task_code":354,"related_runs":355,"delta_psnr":356,"delta_wallclock":357,"rank":34,"verdict":358,"impact_summary":359,"detail_path":360},"Phase H Lego Pareto sweep — stop_iter=10000 で brush dominate、12500 で Phase D dominate (Lego)、ただし scene-dependent 確定","2026-05-26","audit",[14],[346,347,348,349,350,351,352,353],"p1-axis1","phase-h","pareto-sweep","stop-iter","sh-progressive","scene-dependent","lego-detail","calibration","P1 Phase H",[],"**Lego stop_iter=12500 = +0.15 dB** vs Phase D 30k、ficus 同 config = **-4.12 dB の fail**","**Lego stop_iter=12500 = -71% wallclock** vs Phase D 30k","scene-dependent-confirmed-future-work-scene-adaptive","Phase G で G.3 alone 30k = universal Pareto improvement (+0.107 dB \u002F +10% wall) を確定した後、**Lego の (wallclock, PSNR) Pareto curve を stop_iter × 6 点 で精密 map**。\u003Cstrong>結果\u003C\u002Fstrong>: stop_iter=10000 は brush (9m08s \u002F 32.04 dB) を完全 Pareto-dominate (8m42s \u002F 35.931 dB = **同時間で +3.89 dB**)、stop_iter=12500 は Phase D 30k (41m54s \u002F 36.106 dB) を Pareto-dominate (11m48s \u002F 36.259 dB = **-71% wallclock + 0.15 dB**)、stop_iter=17500 以降は marginal +0.1 dB のみで diminishing returns 顕著。\u003Cstrong>anomaly\u003C\u002Fstrong>: stop_iter=25000 が 36.092 dB と 20000 (36.359) より -0.27 dB 低い (variance or training instability の可能性)。\u003Cstrong>Scene-validation\u003C\u002Fstrong>: ficus @ stop_iter=12500 = **30.103 dB** (Phase D 34.22 比 -4.12 dB の大幅 fail)、**Lego sweet spot が scene-dependent と確定**。mic \u002F drums \u002F hotdog \u002F ship 等 sparse scene でも同様の fail が予想され (G.1 stacked パターンと整合)、universal config では Phase G.3 alone 30k が依然 best。\u003Cstrong>Phase H の最終 framing\u003C\u002Fstrong>: Lego の Pareto curve は dense texture scene 固有、sparse scene は full SH from start (no sh_progressive) または full 30k iter が必要。\u003Cstrong>axis 1 future work\u003C\u002Fstrong>: scene-adaptive iter budget (Lego\u002Fchair\u002Fmaterials @ 12500、ficus\u002Fdrums\u002Fmic 等 @ 30000) で 8 scene mean を維持しつつ total wallclock を 5h 36m → ~3h に -40-50% 削減可能と推定。","\u002Ffindings\u002Fp1-axis1-phase-h-lego-pareto-sweep\u002F",{"id":28,"title":362,"date":342,"status":10,"polarity":363,"category":364,"axes":365,"tags":366,"task_code":371,"related_runs":372,"delta_psnr":375,"delta_wallclock":376,"rank":34,"verdict":377,"impact_summary":378,"detail_path":379},"Phase I scene-adaptive iter budget — **STRONG Pareto improvement** 確定 (8 scene mean +0.077 dB at -24% wallclock vs Phase D)","positive","design",[14],[346,367,368,369,349,350,370,353],"phase-i","scene-adaptive","pareto-front","universal-improvement","P1 Phase I",[373,374],"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",{"id":31,"title":381,"date":382,"status":10,"polarity":363,"category":364,"axes":383,"tags":384,"task_code":394,"related_runs":395,"delta_psnr":401,"delta_wallclock":402,"rank":34,"verdict":403,"impact_summary":404,"detail_path":405},"Phase 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)","2026-05-25",[14],[346,385,350,386,369,387,388,389,390,391,392,393],"phase-g","compute-reduction","lego-5k","lego-30k","stacked-config","implementation","unit-tests","bit-exact","smoke-artifact","P1 Phase G.3",[396,397,398,399,400],"lego-phase-g3-sh-progressive-5k","lego-phase-g3-sh-progressive-30k","lego-phase-g1g3-stacked-15k","lego-phase-f1-baseline-5k","lego-brushcompat-opacdecay-30k","+0.15 dB stacked (vs Phase D)、+0.28 dB 30k single (vs Phase D)、-0.12 dB 5k smoke (許容)","**-61% stacked** (vs Phase D)、-1.9% 30k single、-13.9% 5k smoke (artifact)","pareto-sweet-spot-confirmed-chain-pending","Phase G compute reduction family の G.3 (SH-progressive growth) を実装 + bit-exact unit tests (11 件、cargo test 43 件 全 pass) + **3 layer の Lego 結果検証**。\u003Cstrong>(1) 5k smoke\u003C\u002Fstrong>: wallclock -13.9% \u002F splats -22% \u002F PSNR -0.12 dB、cascading splat reduction を観測。\u003Cstrong>(2) 30k full validation\u003C\u002Fstrong>: wallclock **-1.9%** (5k から大幅縮小)、splats **+30%** (5k から逆転)、PSNR **+0.28 dB** (quality improvement!)。5k smoke の cascading 効果は refine.stop_iter=1500 による artifact、30k では sh unlock 完了 (iter 3000) 後に refine が iter 15000 まで full SH で継続 → splats baseline より grow。\u003Cstrong>(3) G.1+G.3 stacked (max_steps=15000 + sh_progressive)\u003C\u002Fstrong>: Lego **16m13s \u002F 36.254 dB \u002F 428k splats** = Phase D 比 **-61% wallclock + 0.15 dB PSNR** で \u003Cstrong>Pareto sweet spot\u003C\u002Fstrong> 確定。\u003Cstrong>Key reframe\u003C\u002Fstrong>: G.3 は「speed win」ではなく「**quality improvement at no speed cost**」、stacked variant で G.1 speed と SH warmup quality gain を統合。\u003Cstrong>Implementation\u003C\u002Fstrong>: \u003Ccode>[trainer.sh_progressive]\u003C\u002Fcode> section (default disabled、全 backward compat)、\u003Ccode>CameraGpu\u003C\u002Fcode> struct を \u003Ccode>sh_degree\u003C\u002Fcode> (buffer layout) と \u003Ccode>active_sh_degree\u003C\u002Fcode> (per-iter eval) に分離。\u003Cstrong>Calibration data point\u003C\u002Fstrong>: Phase F 5 連続 falsification + G.3 5k smoke artifact = audit \u002F smoke overestimate 6 例目、「smoke は production scale を representative しない」が新教訓。8 scene chain validation pending。\u003Cstrong>卒論 narrative\u003C\u002Fstrong>: Phase F (kernel-level fail) → G.2 (architectural insight) → G.3 (algorithmic reframe: speed → quality + stacked Pareto) の 3 family 比較で「Apple Silicon native 最適化は \u003Cstrong>algorithmic compute reduction + early stop の組み合わせが Pareto-optimal\u003C\u002Fstrong>」という構造的 calibration。","\u002Ffindings\u002Fp1-axis1-phase-g3-sh-progressive\u002F",{"id":29,"title":407,"date":382,"status":10,"polarity":363,"category":11,"axes":408,"tags":409,"task_code":420,"related_runs":421,"delta_psnr":429,"delta_wallclock":430,"rank":34,"verdict":431,"impact_summary":432,"detail_path":433},"P1.D multi-scene Phase D re-chain final — 8 scene mean 33.49 dB、brush mean 32.86 を +0.63 dB 上回り、universal win-win-win 実証",[14,15,16],[410,411,412,413,414,415,416,417,418,419],"p1","phase-d","milestone-m5","multi-scene","brush-parity","brush-超え","premultiplied","opacity-decay","universal-win-win-win","rechain-final","P1.D multi-scene re-chain (M5 final)",[400,422,423,424,425,426,427,428],"chair-brushcompat-opacdecay-30k","ficus-brushcompat-opacdecay-30k","drums-brushcompat-opacdecay-30k","hotdog-brushcompat-opacdecay-30k","mic-brushcompat-opacdecay-30k","materials-brushcompat-opacdecay-30k","ship-brushcompat-opacdecay-30k","8 scene mean +0.63 dB vs brush paper (33.49 vs 32.86)","-61% total chain (13h+ → 5h 5m)","accepted-m5-complete","Phase D opacity_decay (rate=0.004 brush default) を 7 scene × 30k full chain bench、Lego val Phase D 30k と合わせて 8 scene 集計。**全 scene で baseline brushcompat 30k 比 PSNR + splats + wallclock すべて改善 (universal win-win-win)**: PSNR +0.18〜+1.42 dB \u002F splats -57〜-78% \u002F wallclock -39〜-69%。8 scene mean 33.49 dB vs brush paper 8 scene mean 32.86 dB = **+0.63 dB 上回り**、本実装が brush の multi-scene mean を decisive に超えた。brush 超え 3 scene (Lego val +4.07 \u002F drums +1.05 \u002F mic +1.02)、4 scene が brush 比 ±0.7 dB 圏内 (chair -0.02 \u002F hotdog -0.39 \u002F ship -0.01 \u002F materials -0.10)、最遠 scene でも ficus -0.65 で接近。全体 wallclock baseline chain (13h+) → Phase D re-chain 5h 5m (-61%)、mean splats 1.4M → 428k (-69%) で brush 282k に肉薄。P1.M5 完全達成 (Lego val > 36 dB ✅ + multi-scene mean > 32 dB ✅)、卒論 central evaluation table の final 数字確定、universal claim 完全実証。","\u002Ffindings\u002Fp1-d-multi-scene-rechain\u002F",1782449788652]