实验定位 负结果但有信息量
A39 接在 A38 后面：既然 TRELLIS.2 raw 能提供更近的 replacement patch，那么最小边界缝合能不能把它接回 A35 的 visible-locked repair surface？本页测试三种三角面级操作：snap_merge、bridge_curtain、snap_bridge。
结果显示：snap_bridge 把平均 boundary edges 从 32095 降到 10123，但 components 从 1596.5 暴涨到 19692.5，nonmanifold edges 从 2791.5 暴涨到 14240。所以它不是可用修复，只是证明“局部距离近”不能自动转化为“拓扑可缝合”。
实验设计（Image-2 风格）
模块设计（Image-2 风格）
Stitching Operators
overlay_control：A38 的未缝合 overlay，对照组。snap_merge：把 patch boundary vertices 吸附到 hole boundary 并合并顶点。bridge_curtain：沿 hole boundary 添加桥接三角面。snap_bridge：先吸附再桥接，最大化边界减少。
Failure Signal
如果 boundary 下降但 components/nonmanifold 暴涨，说明操作是在用局部顶点塌缩制造拓扑病变，而不是真正闭合 surface。这个信号直接支持下一步转向 local implicit/SDF constrained completion。
实验结果（表格）
variant 聚合
| variant | 水密率 | boundary | boundary Δ | components | component Δ | nonmanifold | nonmanifold Δ | faces | surface proxy Chamfer |
|---|---|---|---|---|---|---|---|---|---|
| A38 overlay 对照 | 0.0% | 32095 | 0 | 1596.5 | 0 | 2791.5 | 0 | 709300 | 0.00889 |
| bridge curtain | 0.0% | 25767.5 | -6327.5 | 1712 | 115.5 | 3408.5 | 617 | 719021 | 0.00895 |
| snap + merge | 0.0% | 16720 | -15375 | 16049.5 | 14453 | 11177.5 | 8386 | 709300 | 0.00902 |
| snap + bridge | 0.0% | 10123 | -21972 | 19692.5 | 18096 | 14240 | 11448.5 | 718680 | 0.00897 |
逐 case 输出
| case | variant | 选择 | boundary | boundary Δ | components | component Δ | nonmanifold | nonmanifold Δ | snap rate | bridge rate | OBJ |
|---|---|---|---|---|---|---|---|---|---|---|---|
| gso_002_input4 | A38 overlay 对照 | candidate | 29574 | 0 | 1974 | 0 | 5150 | 0 | - | - | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a39_boundary_aware_patch_stitching/gso_002_input4/mesh_overlay_control_trellis2_raw.obj |
| gso_002_input4 | snap + merge | candidate | 7085 | -22489 | 24262 | 22288 | 17250 | 12100 | 0.9506 | - | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a39_boundary_aware_patch_stitching/gso_002_input4/mesh_snap_merge_trellis2_raw.obj |
| gso_002_input4 | bridge curtain | candidate | 25294 | -4280 | 2052 | 78 | 5647 | 497 | - | 0.9025 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a39_boundary_aware_patch_stitching/gso_002_input4/mesh_bridge_curtain_trellis2_raw.obj |
| gso_002_input4 | snap + bridge | best boundary | 2146 | -27428 | 27129 | 25155 | 19375 | 14225 | 0.9506 | 0.9133 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a39_boundary_aware_patch_stitching/gso_002_input4/mesh_snap_bridge_trellis2_raw.obj |
| gso_008_input4 | A38 overlay 对照 | candidate | 34616 | 0 | 1219 | 0 | 433 | 0 | - | - | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a39_boundary_aware_patch_stitching/gso_008_input4/mesh_overlay_control_trellis2_raw.obj |
| gso_008_input4 | snap + merge | candidate | 26355 | -8261 | 7837 | 6618 | 5105 | 4672 | 0.3784 | - | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a39_boundary_aware_patch_stitching/gso_008_input4/mesh_snap_merge_trellis2_raw.obj |
| gso_008_input4 | bridge curtain | candidate | 26241 | -8375 | 1372 | 153 | 1170 | 737 | - | 0.7611 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a39_boundary_aware_patch_stitching/gso_008_input4/mesh_bridge_curtain_trellis2_raw.obj |
| gso_008_input4 | snap + bridge | best boundary | 18100 | -16516 | 12256 | 11037 | 9105 | 8672 | 0.3784 | 0.8130 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a39_boundary_aware_patch_stitching/gso_008_input4/mesh_snap_bridge_trellis2_raw.obj |
可视化结果
实验结论
- A39 证伪了 naive triangle-level stitching：所有 variant 的 watertight rate 都是 0。
snap_bridge的 boundary 最低，但它通过大量顶点吸附和桥接三角面制造了 component/nonmanifold 爆炸，因此不是有效 mesh repair。bridge_curtain更保守，component 增量较小，但 boundary 降幅有限，仍无法闭合大面积 conflict patch。- 局部替换的真正难点不是 patch 位置，而是 boundary topology correspondence：repair surface 和 raw patch 的边界环数量、顺序和采样密度都不一致。
- 这一步把论文主线进一步收窄：最终方法不应依赖后验三角面 surgery，而应做 visible-locked local implicit completion 或在 TRELLIS.2 生成期约束 observable-state。
下一步想法
- A40：local SDF constrained completion。把 kept visible surface 和 replacement patch 变成局部点/法线约束，只在 conflict tube 内重建等值面。
- 设计 occupancy/signed-distance band：visible-locked 区域作为硬约束，conflict/unobserved 区域允许补全，避免全局体素闭合损失细节。
- 引入 boundary loop matching 作为诊断指标，而不是直接拼接：统计 loop 数量、长度分布、最近邻双向覆盖和曲率兼容性。
- 把 A39 作为论文中的反证实验：mesh 后处理 stitching 不稳定，说明需要生成过程里的 observable-state consistency。