实验定位 小步校准正向 / selector 仍保守
A54 针对 A53 的 saturated conflict case gso_008_input4 做 step calibration。A53 回退 carrier 是 Chamfer-safe，但也完全放弃了 F@5 改善；A54 问的是：如果把 A52 的大步 residual 缩小，是否存在既不伤 Chamfer、又保留部分覆盖收益的区域。
结果：扫描 16 个候选。proxy-safe selector 选 outward_step002，ΔChamfer=-0.000126，ΔF@5=-0.000251；GT Chamfer 最佳是 outward_step006，ΔChamfer=-0.000232；GT F@5 最佳是 inward_counterfactual_step014，但它明显牺牲 Chamfer。
实验设计（Image-2 风格）
模块设计（Image-2 风格）
Calibration Mechanics
- 只对 gso008 计算一次 VGGT depth signed step field。
- 扫描 max_step 0.002/0.004/0.006/0.008/0.014。
- 方向包含 signed、outward-only、inward counterfactual。
- proxy-safe selector 限制 visible/conflict/improved p95 漂移。
Interpretation
A54 证明 A53 的“拒绝大步编辑”不是终点：saturated conflict 仍可能有小步甜点区。但当前 no-GT proxy 过保守，选到了 outward_step002，而 GT Chamfer 最佳是 outward_step006。
实验结果（表格）
| variant | mode | max step | ΔChamfer | ΔF@5 | visible Δ | conflict Δ | improved Δ | depth valid | outward | p95 move | OBJ |
|---|---|---|---|---|---|---|---|---|---|---|---|
| carrier_control | carrier | 0.000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.0% | 0.0% | 0.00000 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a54_saturated_conflict_step_calibration/gso_008_input4/mesh_carrier_control.obj |
| inward_counterfactual_step002 | inward_counterfactual | 0.002 | -0.000053 | 0.001771 | 0.000239 | 0.000334 | 0.000109 | 84.0% | 0.0% | 0.00200 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a54_saturated_conflict_step_calibration/gso_008_input4/mesh_inward_counterfactual_step002.obj |
| inward_counterfactual_step004 | inward_counterfactual | 0.004 | 0.000198 | 0.003717 | 0.000502 | 0.000853 | 0.000377 | 84.0% | 0.0% | 0.00400 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a54_saturated_conflict_step_calibration/gso_008_input4/mesh_inward_counterfactual_step004.obj |
| inward_counterfactual_step006 | inward_counterfactual | 0.006 | 0.000636 | 0.005814 | 0.000875 | 0.001512 | 0.001087 | 84.0% | 0.0% | 0.00600 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a54_saturated_conflict_step_calibration/gso_008_input4/mesh_inward_counterfactual_step006.obj |
| inward_counterfactual_step008 | inward_counterfactual | 0.008 | 0.001743 | 0.006825 | 0.001354 | 0.002314 | 0.000790 | 84.0% | 0.0% | 0.00800 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a54_saturated_conflict_step_calibration/gso_008_input4/mesh_inward_counterfactual_step008.obj |
| inward_counterfactual_step014 | inward_counterfactual | 0.014 | 0.004461 | 0.010752 | 0.003475 | 0.006263 | 0.003516 | 84.0% | 0.0% | 0.01400 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a54_saturated_conflict_step_calibration/gso_008_input4/mesh_inward_counterfactual_step014.obj |
| outward_step002 | outward | 0.002 | -0.000126 | -0.000251 | 0.000008 | 0.000077 | -0.000068 | 84.0% | 53.4% | 0.00200 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a54_saturated_conflict_step_calibration/gso_008_input4/mesh_outward_step002.obj |
| outward_step004 | outward | 0.004 | 0.000395 | -0.001951 | 0.000167 | 0.000566 | 0.000140 | 84.0% | 53.4% | 0.00400 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a54_saturated_conflict_step_calibration/gso_008_input4/mesh_outward_step004.obj |
| outward_step006 | outward | 0.006 | -0.000232 | 0.000952 | 0.000537 | 0.001212 | 0.000334 | 84.0% | 53.4% | 0.00600 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a54_saturated_conflict_step_calibration/gso_008_input4/mesh_outward_step006.obj |
| outward_step008 | outward | 0.008 | 0.000934 | 0.000191 | 0.000921 | 0.002245 | 0.000715 | 84.0% | 53.4% | 0.00800 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a54_saturated_conflict_step_calibration/gso_008_input4/mesh_outward_step008.obj |
| outward_step014 | outward | 0.014 | 0.001474 | 0.003920 | 0.003262 | 0.005800 | 0.002678 | 84.0% | 53.4% | 0.01400 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a54_saturated_conflict_step_calibration/gso_008_input4/mesh_outward_step014.obj |
| signed_step002 | signed | 0.002 | 0.000276 | -0.000441 | 0.000367 | 0.000378 | 0.000013 | 84.0% | 53.4% | 0.00200 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a54_saturated_conflict_step_calibration/gso_008_input4/mesh_signed_step002.obj |
| signed_step004 | signed | 0.004 | 0.000985 | 0.001593 | 0.000838 | 0.001176 | 0.000771 | 84.0% | 53.4% | 0.00400 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a54_saturated_conflict_step_calibration/gso_008_input4/mesh_signed_step004.obj |
| signed_step006 | signed | 0.006 | 0.002441 | 0.001253 | 0.001616 | 0.002667 | 0.001400 | 84.0% | 53.4% | 0.00600 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a54_saturated_conflict_step_calibration/gso_008_input4/mesh_signed_step006.obj |
| signed_step008 | signed | 0.008 | 0.003211 | 0.002940 | 0.002727 | 0.003916 | 0.001950 | 84.0% | 53.4% | 0.00800 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a54_saturated_conflict_step_calibration/gso_008_input4/mesh_signed_step008.obj |
| signed_step014 | signed | 0.014 | 0.005894 | 0.008851 | 0.006191 | 0.008207 | 0.005376 | 84.0% | 53.4% | 0.01400 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a54_saturated_conflict_step_calibration/gso_008_input4/mesh_signed_step014.obj |
可视化结果
实验结论
- A54 证实 gso008 不是只能 carrier-only：
outward_step006达到 ΔChamfer=-0.000232，且 F@5 也小幅正向。 - A52 的大步 outward ΔChamfer=0.001081，A54 小步 outward 把这个反向修成正向，说明问题主要是 step scale，而不只是 direction。
- proxy-safe selector 选
outward_step002，ΔChamfer=-0.000126，但 ΔF@5=-0.000251 微负；它比 A53 carrier 更有几何收益，但仍低估了 outward_step006 的甜点区。 - F@5 最强的 inward/signed 大步仍然伤 Chamfer，说明 F@5 覆盖收益和形状保真存在真实冲突，不能单指标选择。
- 论文价值：A54 把 A53 的 reject option 推进为 calibrated residual，不是简单“拒绝编辑”，而是先拒绝大步，再搜索可置信小步。
下一步想法
- A55：学习/拟合 no-GT step-size selector，让 proxy 从 outward_step002 逼近 outward_step006，而不是只做硬阈值。
- A56：把 A54 的 calibrated small-step 合并回 10-case pipeline，替换 A53 对 gso008 的 carrier fallback。
- A57：加入 20-view render metrics，确认 outward_step006 的 Chamfer 正向是否也表现为更好的法线/轮廓/LPIPS。
- A58：对其他 high-conflict 或 not-ready case，尤其 gso006，先修 carrier topology 后再跑 small-step calibration。