实验定位 局部性有效 / 几何仍负
A45 是对 A44 失败模式的直接修正：A44 的 active vertex ratio 接近 100%，说明它不是局部编辑，而是在全局把 carrier 拉向观测点云。A45 用 A35 的 surface_conflict face mask 定义活动区，用非冲突 visible surface 定义 hard lock，只允许 conflict band 内的顶点接收 replacement patch residual。
结果表明 locality gate 起作用：local_tau035_lock025 active vertex ratio=31.7%、hard lock=40.9%，远低于 A44 的全局移动；同时 topology 仍保持 watertight=100.0%、components=1。但 GT Chamfer 从 carrier 的 0.08836 变到 0.08897，仍然变差。
实验设计（Image-2 风格）
模块设计（Image-2 风格）
Local Activity Gate
- activity band：carrier vertex 到 A35 conflict samples 的距离。
- hard lock：靠近非冲突 visible surface 且不在 conflict core 的顶点不动。
- residual target：A38 replacement patch samples。
- 拓扑 invariant：不改 connectivity，只移动顶点。
Why It Matters
A45 把 A44 的“signed carrier”从全局 attraction 改成局部 residual，测试 observable-state edit 的关键条件：只动坏区域、锁住可信可见面、保持水密拓扑。
实验结果（表格）
聚合结果
| 变体 | GT Chamfer↓ | F@5↑ | 水密率↑ | components↓ | boundary↓ | nonmanifold↓ | visible p95↓ | conflict p95↓ | patch p95↓ | active vertices↓ | hard lock↑ | 结论 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| A42 carrier control | 0.08836 | 0.4551 | 100.0% | 1 | 0 | 0 | 0.03233 | 0.03721 | 0.03527 | 0.0% | 0.0% | carrier baseline |
| local tau035 lock025 | 0.08897 | 0.4512 | 100.0% | 1 | 0 | 0 | 0.03035 | 0.03443 | 0.03341 | 31.7% | 40.9% | 局部性有效 / GT 仍差 |
| local tau050 lock030 | 0.08931 | 0.4488 | 100.0% | 1 | 0 | 0 | 0.03086 | 0.03415 | 0.03329 | 43.4% | 51.8% | 局部性有效 / GT 仍差 |
| local tau065 soft | 0.08886 | 0.4528 | 100.0% | 1 | 0 | 0 | 0.03221 | 0.03576 | 0.03396 | 39.6% | 58.9% | 局部性有效 / GT 仍差 |
| local tau050 smooth | 0.08934 | 0.4514 | 100.0% | 1 | 0 | 0 | 0.03150 | 0.03517 | 0.03395 | 43.4% | 51.8% | 局部性有效 / GT 仍差 |
逐 case 输出
| case | 变体 | GT Chamfer↓ | F@5↑ | watertight | components | visible p95 | conflict p95 | active vertices | hard lock | p95 move | OBJ |
|---|---|---|---|---|---|---|---|---|---|---|---|
| gso_002_input4 | A42 carrier control | 0.04541 | 0.6927 | True | 1 | 0.03327 | 0.03480 | 0.0% | 0.0% | 0.00000 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a45_conflict_local_signed_residual/gso_002_input4/mesh_carrier_control.obj |
| gso_002_input4 | local tau035 lock025 | 0.04537 | 0.6927 | True | 1 | 0.03197 | 0.03194 | 16.7% | 59.9% | 0.00810 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a45_conflict_local_signed_residual/gso_002_input4/mesh_local_tau035_lock025.obj |
| gso_002_input4 | local tau050 lock030 | 0.04533 | 0.6901 | True | 1 | 0.03201 | 0.03295 | 18.4% | 72.4% | 0.00720 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a45_conflict_local_signed_residual/gso_002_input4/mesh_local_tau050_lock030.obj |
| gso_002_input4 | local tau065 soft | 0.04557 | 0.6916 | True | 1 | 0.03307 | 0.03436 | 17.0% | 80.1% | 0.00420 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a45_conflict_local_signed_residual/gso_002_input4/mesh_local_tau065_lock035_soft.obj |
| gso_002_input4 | local tau050 smooth | 0.04552 | 0.6928 | True | 1 | 0.03247 | 0.03350 | 18.4% | 72.4% | 0.00490 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a45_conflict_local_signed_residual/gso_002_input4/mesh_local_tau050_lock030_smooth.obj |
| gso_008_input4 | A42 carrier control | 0.13132 | 0.2176 | True | 1 | 0.03138 | 0.03962 | 0.0% | 0.0% | 0.00000 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a45_conflict_local_signed_residual/gso_008_input4/mesh_carrier_control.obj |
| gso_008_input4 | local tau035 lock025 | 0.13257 | 0.2098 | True | 1 | 0.02873 | 0.03693 | 46.7% | 21.8% | 0.00810 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a45_conflict_local_signed_residual/gso_008_input4/mesh_local_tau035_lock025.obj |
| gso_008_input4 | local tau050 lock030 | 0.13329 | 0.2075 | True | 1 | 0.02970 | 0.03536 | 68.5% | 31.3% | 0.00720 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a45_conflict_local_signed_residual/gso_008_input4/mesh_local_tau050_lock030.obj |
| gso_008_input4 | local tau065 soft | 0.13214 | 0.2139 | True | 1 | 0.03136 | 0.03715 | 62.3% | 37.7% | 0.00420 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a45_conflict_local_signed_residual/gso_008_input4/mesh_local_tau065_lock035_soft.obj |
| gso_008_input4 | local tau050 smooth | 0.13315 | 0.2099 | True | 1 | 0.03054 | 0.03684 | 68.5% | 31.3% | 0.00490 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a45_conflict_local_signed_residual/gso_008_input4/mesh_local_tau050_lock030_smooth.obj |
可视化结果
实验结论
- A45 成功修掉 A44 的一个明确问题：active vertex ratio 从接近 100% 降到 31.7% - 43.4%，说明 conflict-local activity gate 是有效机制。
- A45 仍保持拓扑 invariant：所有局部 residual 变体 watertight=100%、boundary=0、nonmanifold=0、components=1。
- 但 GT 几何仍未改善：最佳非 control 的
local_tau065_lock035_softChamfer=0.08886，仍差于 carrier control 的 0.08836。 - 这说明 A44 的失败不只来自“动太多”，还来自“移动方向不可靠”：nearest replacement patch attraction 会改善 conflict proxy，却可能偏离真实 GT。
- 下一步不能继续调 tau/alpha；应把 residual direction 改成 signed depth residual / normal-consistent displacement，并只在 conflict-local tube 内解一个局部优化问题。
下一步想法
- A46：Depth-signed displacement，用 VGGT/candidate depth residual 决定沿法线内外移动，而不是最近 patch 点吸引。
- A47：Local optimization with hard constraints，把 visible-lock 顶点作为 Dirichlet boundary，conflict 顶点作为变量，优化 depth residual + Laplacian smooth。
- A48：加入 residual direction sanity check：如果 patch attraction 与 carrier normal / VGGT depth sign 冲突，则 defer 而不是移动。
- A49：把 A45/A46 的 active ratio、hard lock ratio、GT delta 做成论文 ablation，证明“局部性”和“有符号方向”是两个独立必要条件。