实验定位 诊断有效 / 指向新 gate
A5 回答 A4 留下的问题：为什么 A3 的 sparse support gate 通过后，完整 TRELLIS.2 解码仍然失败。关键发现是：gso_000_input4 在 selected alpha 下 sparse support 与 baseline 精确一致，但 full decode 仍报 empty tensor。这把问题从 sparse support 阶段推进到了 shape/SLat/decoder viability 阶段。
实验设计（Image-2 风格）
模块设计（Image-2 风格）
本页的模块设计图已在“实验设计”区域并排展示；它描述输入视角、VGGT/TRELLIS.2 信号、门控/路由、mesh 输出和评测反馈之间的数据流。这里单独标出模块设计章节，方便按统一汇报格式阅读。
实验结果（表格）
| case | selected alpha | sparse match | sparse Jaccard | decode 状态 | 风险类型 | Chamfer Δ ↓ | Watertight | 解释 |
|---|---|---|---|---|---|---|---|---|
| gso_000_input4 | 0.1250 | 精确一致 | 1.000 | failed | decode empty | - | - | sparse support gate passed, but full decode failed before mesh extraction |
| gso_001_input4 | 1.0000 | 近似/漂移 | 0.993 | ok | 拓扑开口 | -0.00211 | 0.0 | full decode produced a mesh, but raw topology stayed open |
| gso_002_input4 | 1.0000 | 近似/漂移 | 0.834 | ok | 拓扑开口 | -0.00014 | 0.0 | full decode produced a mesh, but raw topology stayed open |
| gso_003_input4 | 1.0000 | 近似/漂移 | 0.952 | ok | 拓扑开口 | -0.00001 | 0.0 | full decode produced a mesh, but raw topology stayed open |
| gso_004_input4 | 1.0000 | 精确一致 | 1.000 | ok | 拓扑开口 | 0.00000 | 0.0 | full decode produced a mesh, but raw topology stayed open |
| gso_005_input4 | 1.0000 | 精确一致 | 1.000 | ok | 拓扑开口 | 0.00000 | 0.0 | full decode produced a mesh, but raw topology stayed open |
| gso_006_input4 | 1.0000 | 近似/漂移 | 0.971 | ok | 几何恶化+开口 | 0.00070 | 0.0 | full decode produced a mesh, but geometry worsened and raw topology stayed open |
| gso_007_input4 | 1.0000 | 近似/漂移 | 0.986 | ok | 几何恶化+开口 | 0.00173 | 0.0 | full decode produced a mesh, but geometry worsened and raw topology stayed open |
| gso_008_input4 | 1.0000 | 近似/漂移 | 0.968 | ok | 几何恶化+开口 | 0.00577 | 0.0 | full decode produced a mesh, but geometry worsened and raw topology stayed open |
| gso_009_input4 | 1.0000 | 近似/漂移 | 0.839 | ok | 几何恶化+开口 | 0.00315 | 0.0 | full decode produced a mesh, but geometry worsened and raw topology stayed open |
可视化结果
所有 case 都通过 A3 的 selected-alpha sparse gate。
gso_000_input4 是 sparse 精确一致但 decode 失败的反例。
可解码的 selected-alpha mesh 仍全部非水密。
轨迹证据
| case | 初始噪声一致 | latent 首次分歧 step | occupancy 首次分歧 step | final sparse Jaccard | uniform tokens | adapter tokens | condition delta RMS |
|---|---|---|---|---|---|---|---|
| gso_000_input4 | True | 1 | 6 | 0.000 | 26 | 24 | 0.00015159 |
trajectory 证据来自单例 step25 诊断：极小 residual 已经能让 latent 在 step 1 分歧、decoded occupancy 在 step 6 分歧；而 A3 selected alpha 又证明最终 sparse support 可以重新对齐。两者合起来说明：只看最终 sparse 坐标会漏掉 decoder 轨迹内部的坍塌风险。
实验结论
- A5 confirms a concrete sparse-to-decoder gap: gso_000_input4 has an exact selected sparse support match, yet full decode fails with an empty-tensor error.
- Sparse token ratio, Jaccard, centroid drift, and exact support equality are necessary safety checks, but they are not sufficient after the shape/decoder stage begins.
- The next paper-worthy direction is a decoder-aware occupancy or SLat viability gate that rejects a VGGT residual before full mesh extraction if the decoded band collapses, explodes, or drifts from the baseline.
结论要保守理解：A5 的“sparse 精确一致”来自 A3 preflight 记录，不等价于 A4 full-run 内部的 sparse support 也一致。后续 A6 已经进一步修正根因：selected-alpha 在真实 full-run 中会先坍塌为空 sparse coords，然后才在 shape SLat 入口崩溃。
限制和不确定性
- A5 is a diagnostic aggregation over existing A3/A4 outputs, not a newly trained model.
- Full shape-SLat occupancy was not captured for all 10 cases, so the empty decode root cause is inferred from sparse equality plus TRELLIS.2 runtime error.
- The experiment uses protocol4_limit10; the gate still needs validation on the 100-case 2/4/6/8 matrix.
下一步想法
- Instrument TRELLIS.2 shape-SLat sampling to capture active token count, decoded occupancy fraction, centroid, and extent before mesh extraction.
- Define a two-stage gate: sparse support gate first, then decoder occupancy viability gate.
- Route baseline, selected-alpha, and repair candidates through a no-extra-failure selector before larger-scale evaluation.