ReSplat Research Roadmap

Renderer Score Tokens:实验路线图

Claim-driven validation、stop/go gates 与 GPU 预算;尚未运行实验

Published: 2026-07-16 Scope: claim-driven validation Status: planning only;所有 runs 均未启动

Experiment Plan

Problem:ReSplat 的 residual 与当前 renderer primitive/action 不对齐,容易在遮挡、重叠和参数耦合区域产生错误更新。 Method Thesis:standard render 后的 analytic score-replay primitive tokens,比 scalar attribution、index-bound residual 更适合作为 recurrent Gaussian updater 的固定维度 action interface,并应接近 gradient-conditioned quality、获得更好的 latency/VRAM Pareto。 Date:2026-07-15 Execution status:PLANNING ONLY;本轮没有启动任何 run。

Claim Map

ClaimWhy It MattersMinimum Convincing EvidenceLinked Blocks
C1:typed renderer score 更能预测有用 Gaussian action这是超越 PRIMU/GaussianPOP 用途迁移质疑的核心相对 index residual、alphaT scatter、GaussianPOP score,held-out one-step utility 更高、loss-increasing step fraction 更低,并在 overlap/occlusion 分层中保持优势B0, B1, B3
C2:score-replay interface 在 matched time/capacity 下改善 NVS证明方法不是只在诊断指标上好看DL3DV 8-view 四步 PSNR 约 +0.25 dB 或遮挡区域显著改善且 global 不退化;端到端 overhead ≤20%;3 seedsB2, B3, B4
Anti-claim:收益不是额外参数、手写近似 gradient 或更多 render否则 paper 会被判为工程迁移matched-capacity projection、raw/detached gradient updater、显式 gradient descent、相同 render/time 对照B3, B4

Paper Storyline

Experiment Blocks

Block 0:Renderer Score Correctness and Identifiability

Block 1:Held-Out Single-Step Action Utility

Block 2:Main Matched-Capacity NVS Result

Block 3:Novelty Isolation and Deletion Study

Block 4:Score-Replay Kernel Efficiency and Scaling

Block 5:Failure Analysis

Run Order and Milestones

MilestoneGoalRunsDecision GateCost EstimateRisk
M0score 公式、suffix recovery 与 replay kernel 正确R000–R004reference error <1e-3;geometry rank 可用;overhead <30%0–20 H200h + 40–100 A6000hCUDA/公式错误
M1official baseline 与 matched baselineR010–R012指标接近论文/官方 checkpoint;seed variance 可接受120–300 H200h基线复现偏差
M2single-step utility 决策R020–R026proposed 超过 alphaT/GaussianPOP;否则 STOP70–180 H200hnovelty 被 scalar baseline 吞掉
M3主方法 3 seedsR030–R032+0.25 dB 或强区域收益且 global 不退化450–900 H200hutility 不转化
M4decisive ablations + gradient kill testR040–R049typed/sign/lift 必须;matched-time gradient 不支配450–1,000 H200h手写 gradient objection
M5full-res、cross-dataset、efficiencyR050–R057overhead ≤20%;跨分辨率/数据集不崩350–900 H200hfused kernel 不扩展
M6failure/qualitative polishR060–R064形成可信边界和可视化100–300 H200h时间不足

Stop / Go Gates

  1. Gate G0 — renderer correctness:M0 不通过,停止。
  2. Gate G1 — novelty isolation:alphaT/GaussianPOP 与 proposed 单步 utility 相同,停止。
  3. Gate G2 — action-to-NVS transfer:四步没有 global 或高难区域收益,停止或降为 workshop。
  4. Gate G3 — backward-free value:matched-wall-time gradient updater 更好且资源差异很小,停止主 claim。
  5. Gate G4 — venue readiness:3 seeds、full-res、效率和 failure analysis 均成立后,才进入投稿写作。

Compute and Data Budget

Timeline

目标
1–2reference score、synthetic diagnostic、数值核验
3–5score-replay prototype、M0 gate
6–8baseline 与 single-step utility、G1 gate
9–12主训练单 seed、G2 gate
13–163 seeds、核心 deletion、gradient kill test
17–19512×960、view-count、RE10K、效率
20–22failure analysis、图表、增量查新
23–24buffer:重跑、写作、审稿风险修补

Risks and Mitigations

Final Checklist