实验定位 方向诊断有效 / 几何负
A46 接着 A45 做 direction ablation：A45 已证明 conflict-local activity gate 有效，但 nearest patch vector 仍伤 GT。A46 保持同样的局部活动区和 visible hard-lock，只把位移改为 carrier vertex normal 上的 signed scalar，分别测试 signed、inward-only、outward-only 和 smooth。
结果把失败原因又收窄了一层：normal signed soft 的 conflict p95 从 carrier 的 0.03746 降到 0.03254，但 GT Chamfer 从 0.08858 变差到 0.08969；normal outward only 几乎不动，Chamfer=0.08865 接近 carrier。最关键的是 active 区域 outward ratio 只有 3.5%，说明 patch residual 大部分指向 inward shrink。
实验设计（Image-2 风格）
模块设计（Image-2 风格）
Signed Normal Residual
- 活动区：沿用 A45 conflict band + visible hard-lock。
- 方向：nearest patch vector 投影到 carrier normal。
- 变体：signed、strong、outward-only、inward-only、smooth。
- 不使用 GT 选择；GT 只用于 retrospective validation。
Interpretation
如果 normal-only 比 A45 好，说明切向 patch attraction 是主要问题；如果仍然变差，则需要真正的 depth sign / normal consistency 来决定方向，而不是从 patch nearest neighbor 推断方向。
实验结果（表格）
聚合结果
| 变体 | GT Chamfer↓ | F@5↑ | 水密率↑ | components↓ | visible p95↓ | conflict p95↓ | patch p95↓ | active vertices | hard lock | outward ratio | p95 move | 结论 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| A42 carrier control | 0.08858 | 0.4523 | 100.0% | 1 | 0.03238 | 0.03746 | 0.03533 | 0.0% | 0.0% | 0.0% | 0.00000 | carrier baseline |
| normal signed soft | 0.08969 | 0.4452 | 100.0% | 1 | 0.03015 | 0.03254 | 0.03298 | 43.4% | 51.8% | 3.5% | 0.01059 | 方向诊断 / GT 未过 |
| normal signed strong | 0.08977 | 0.4441 | 100.0% | 1 | 0.02906 | 0.03156 | 0.03252 | 43.4% | 51.8% | 3.5% | 0.01600 | proxy 好但 GT 差 |
| normal outward only | 0.08865 | 0.4504 | 100.0% | 1 | 0.03242 | 0.03720 | 0.03554 | 43.4% | 51.8% | 3.5% | 0.00000 | 方向诊断 / GT 未过 |
| normal inward only | 0.09018 | 0.4423 | 100.0% | 1 | 0.02895 | 0.03302 | 0.03241 | 43.4% | 51.8% | 3.5% | 0.01600 | proxy 好但 GT 差 |
| normal signed smooth | 0.08965 | 0.4445 | 100.0% | 1 | 0.02997 | 0.03251 | 0.03300 | 43.4% | 51.8% | 3.5% | 0.01198 | 方向诊断 / GT 未过 |
逐 case 输出
| case | 变体 | GT Chamfer↓ | F@5↑ | watertight | components | conflict p95 | outward ratio | p95 signed step | tangential residual | OBJ |
|---|---|---|---|---|---|---|---|---|---|---|
| gso_002_input4 | A42 carrier control | 0.04569 | 0.6914 | True | 1 | 0.03539 | - | - | - | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a46_signed_direction_residual/gso_002_input4/mesh_carrier_control.obj |
| gso_002_input4 | normal signed soft | 0.04560 | 0.6851 | True | 1 | 0.03142 | 6.1% | 0.01200 | 0.01448 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a46_signed_direction_residual/gso_002_input4/mesh_normal_signed_soft.obj |
| gso_002_input4 | normal signed strong | 0.04530 | 0.6899 | True | 1 | 0.02973 | 6.1% | 0.01600 | 0.01448 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a46_signed_direction_residual/gso_002_input4/mesh_normal_signed_strong.obj |
| gso_002_input4 | normal outward only | 0.04562 | 0.6897 | True | 1 | 0.03495 | 6.1% | 0.00658 | 0.01448 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a46_signed_direction_residual/gso_002_input4/mesh_normal_outward_only.obj |
| gso_002_input4 | normal inward only | 0.04544 | 0.6874 | True | 1 | 0.03100 | 6.1% | 0.01600 | 0.01448 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a46_signed_direction_residual/gso_002_input4/mesh_normal_inward_only.obj |
| gso_002_input4 | normal signed smooth | 0.04544 | 0.6876 | True | 1 | 0.03174 | 6.1% | 0.01200 | 0.01448 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a46_signed_direction_residual/gso_002_input4/mesh_normal_signed_smooth.obj |
| gso_008_input4 | A42 carrier control | 0.13147 | 0.2132 | True | 1 | 0.03953 | - | - | - | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a46_signed_direction_residual/gso_008_input4/mesh_carrier_control.obj |
| gso_008_input4 | normal signed soft | 0.13378 | 0.2054 | True | 1 | 0.03365 | 0.9% | 0.01097 | 0.00741 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a46_signed_direction_residual/gso_008_input4/mesh_normal_signed_soft.obj |
| gso_008_input4 | normal signed strong | 0.13425 | 0.1983 | True | 1 | 0.03338 | 0.9% | 0.01600 | 0.00741 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a46_signed_direction_residual/gso_008_input4/mesh_normal_signed_strong.obj |
| gso_008_input4 | normal outward only | 0.13168 | 0.2111 | True | 1 | 0.03945 | 0.9% | 0.00000 | 0.00741 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a46_signed_direction_residual/gso_008_input4/mesh_normal_outward_only.obj |
| gso_008_input4 | normal inward only | 0.13491 | 0.1972 | True | 1 | 0.03505 | 0.9% | 0.01600 | 0.00741 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a46_signed_direction_residual/gso_008_input4/mesh_normal_inward_only.obj |
| gso_008_input4 | normal signed smooth | 0.13386 | 0.2014 | True | 1 | 0.03327 | 0.9% | 0.01200 | 0.00741 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a46_signed_direction_residual/gso_008_input4/mesh_normal_signed_smooth.obj |
可视化结果
实验结论
- A46 证明“去掉切向 patch attraction”仍不够：normal signed soft/strong 都改善 conflict proxy，但 GT Chamfer 和 F@5 变差。
- normal inward-only 的 proxy 最好之一，conflict p95=0.03302，但 Chamfer=0.09018 最差，说明 inward shrink 是危险方向。
- outward-only 几乎不动，GT 最接近 carrier；结合 outward ratio=3.5%，说明 replacement patch 对当前 carrier 的 signed signal 绝大多数是 inward。
- 这给出更深的论文结论：observable-state guidance 不能从 nearest replacement patch 自动推出正确 signed direction，必须引入 view depth residual、normal agreement 或可见轮廓约束来判断内外。
- A44/A45/A46 已形成三段必要条件：保拓扑 carrier、conflict-local activity、有符号方向；其中前两个可行，第三个仍缺少可靠观测符号。
下一步想法
- A47：Depth-sign oracle/proxy，对 active vertices 投影到输入视角，用 VGGT/candidate depth residual 决定 inward/outward，而不是 patch nearest neighbor。
- A48：Direction sanity gate，如果 patch-derived signed direction 与 depth-derived sign 冲突，则 defer 或只做 smoothing。
- A49：Normal agreement filter，只允许 patch normal 与 carrier normal 一致的局部 residual 生效。
- A50：把 A44-A46 写成论文 ablation：topology carrier、local activity、signed direction 三者缺一不可。