实验定位 诊断正向 / 非最终 mesh
A38 不是再尝试“删除后补洞”，而是问一个更像论文问题的点：A35 标出的 conflict surface 能否从已有生成候选中找到可信替代表面？如果可以，后续就能做局部 patch replacement 或 SDF constrained completion；如果找不到，就说明要回到生成期约束。
结果很清楚：TRELLIS.2 raw 在两个 case 都是 best proxy source，mean coverage@2%=65.7%，mean normal |cos|=0.7622，boundary gap p95=0.07644；而 TRELLIS raw 的 coverage@2%=26.9%，normal |cos|=0.5131。这说明替代表面更可能来自同一个 TRELLIS.2 raw trajectory，而不是跨模型拼接。
实验设计（Image-2 风格）
模块设计（Image-2 风格）
Patch Proxy Operator
- 输入：A35 face labels 与 repair mesh。
- source：TRELLIS.2 raw 和 TRELLIS raw 两类候选表面。
- 映射：conflict face centroid 到 source face centroid 的近邻距离。
- 选择：半径阈值 + 法线兼容性得到 source patch。
- 输出：
patch_*.obj和未缝合的hybrid_overlay_*.obj。
Why It Matters
A38 把“该删哪里”推进到“用哪里替换”。如果 conflict 区域在 raw baseline 中有近邻且法线兼容，后续可以做边界感知 stitching 或局部 SDF completion；如果 overlay 仍非水密，就说明问题不是找 patch，而是如何受约束地缝合。
实验结果（表格）
source 聚合
| source | mean dist | p95 dist | coverage@2% | coverage@5% | normal |cos| | boundary gap | boundary p95 | patch faces | overlay boundary | overlay components | overlay watertight |
|---|---|---|---|---|---|---|---|---|---|---|---|
| TRELLIS.2 raw | 0.02554 | 0.08598 | 65.7% | 79.7% | 0.7622 | 0.02250 | 0.07644 | 488630 | 32095 | 1596.5 | 0.0% |
| TRELLIS raw | 0.05198 | 0.12703 | 26.9% | 51.0% | 0.5131 | 0.07290 | 0.18317 | 181056 | 22858 | 453.5 | 0.0% |
逐 case 输出
| case | source | 选择 | conflict ratio | mean dist | p95 dist | coverage@2% | coverage@5% | normal |cos| | boundary p95 | patch faces | overlay boundary | overlay components | patch OBJ | overlay OBJ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| gso_002_input4 | TRELLIS.2 raw | best | 0.1972 | 0.03758 | 0.10196 | 47.0% | 68.7% | 0.7026 | 0.13563 | 778187 | 29574 | 1974 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a38_baseline_nearest_patch_proxy/gso_002_input4/patch_trellis2_raw.obj | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a38_baseline_nearest_patch_proxy/gso_002_input4/hybrid_overlay_trellis2_raw.obj |
| gso_002_input4 | TRELLIS raw | candidate | 0.1972 | 0.04910 | 0.10250 | 26.8% | 50.5% | 0.6819 | 0.14756 | 324925 | 29198 | 529 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a38_baseline_nearest_patch_proxy/gso_002_input4/patch_trellis_raw.obj | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a38_baseline_nearest_patch_proxy/gso_002_input4/hybrid_overlay_trellis_raw.obj |
| gso_008_input4 | TRELLIS.2 raw | best | 0.6948 | 0.01351 | 0.06999 | 84.5% | 90.6% | 0.8217 | 0.01726 | 199073 | 34616 | 1219 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a38_baseline_nearest_patch_proxy/gso_008_input4/patch_trellis2_raw.obj | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a38_baseline_nearest_patch_proxy/gso_008_input4/hybrid_overlay_trellis2_raw.obj |
| gso_008_input4 | TRELLIS raw | candidate | 0.6948 | 0.05487 | 0.15156 | 27.1% | 51.4% | 0.3444 | 0.21877 | 37188 | 16518 | 378 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a38_baseline_nearest_patch_proxy/gso_008_input4/patch_trellis_raw.obj | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a38_baseline_nearest_patch_proxy/gso_008_input4/hybrid_overlay_trellis_raw.obj |
可视化结果
实验结论
- A38 给出一个正向机制证据：conflict patch 不是完全无解，TRELLIS.2 raw 通常在局部有更近、更法线兼容的替代表面。
- TRELLIS.2 raw 明显优于 TRELLIS raw：coverage@2%、coverage@5%、normal |cos| 和 boundary gap 都更好，说明局部替换应优先在同一生成族或同一 trajectory 的候选之间做。
- 但 A38 也给出明确负面边界：
hybrid_overlay不是 stitched final mesh，overlay watertight rate=0.0%，boundary/components 仍然很高。 - 因此现在瓶颈从“找不到替换表面”转移到了“如何把替换 patch 以拓扑一致方式缝回 visible-locked surface”。
- 这比工程式 A+B 更接近论文贡献：observable-state 不是简单筛选器，而是定义了局部替换约束、边界约束和生成期补全域。
下一步想法
- A39：boundary-aware patch stitching。沿 hole boundary 建立 repair boundary vertex 到 source patch boundary/nearest surface 的对应，并尝试局部 bridge faces 或 constrained remeshing。
- 更稳的路线是局部 SDF completion：把 visible-locked surface 与 replacement patch 作为带符号/无符号约束，只在 conflict/unobserved tube 内重建等值面。
- 把 A38 的 proxy metrics 变成自动路由器：coverage、normal compatibility、boundary gap 三项决定是否使用 raw patch replacement。
- 把 patch source 从两个固定 mesh 扩展为 candidate pool：TRELLIS.2 多 seed、repair candidate、VGGT-guided candidate，寻找最小 boundary gap 的局部 surface。