实验定位 机制正向 / 非最终高保真
A40 从 A39 的三角面 stitching 转向隐式补全:把 A35 的 visible-kept surface 与 A38 的 TRELLIS.2 raw replacement patch 采样成点云,在局部包围盒中构造 unsigned distance field envelope,再用 marching cubes 导出 mesh。
最佳轻量原型 udf_pitch025_r035 在两个 case 上 watertight rate=100.0%、mean boundary=0、mean nonmanifold=0、mean components=1,相对 A39 snap_bridge 的 component 平均减少 19691.5。代价是 visible p95 drift=0.04150,说明 UDF envelope 会厚化/平滑表面。
实验设计(Image-2 风格)
模块设计(Image-2 风格)
Implicit Proxy
- 输入:A35 non-conflict kept surface + A38 best replacement patch。
- 采样:每个 case 约 52k surface points。
- 距离场:KDTree 最近表面距离,形成 UDF。
- 重建:在
distance = radius处 marching cubes。 - 对照:相对 A39
snap_bridge看 boundary/component/nonmanifold 是否下降。
Unexpected Issue
Open3D Poisson depth8 原型超过交互运行时间并被停止;A40 改用轻量 UDF marching-cubes。这个选择让实验可复现,但也意味着它是 topology feasibility proxy,不是最终 SDF 高保真补全。
实验结果(表格)
variant 聚合
| variant | 水密率 | boundary | components | nonmanifold | faces | proxy Chamfer | visible p95 | patch p95 | boundary Δ vs A39 | component Δ vs A39 | nonmanifold Δ vs A39 |
|---|---|---|---|---|---|---|---|---|---|---|---|
| UDF p=.025 r=.035 | 100.0% | 0 | 1 | 0 | 18316 | 0.03524 | 0.04150 | 0.04334 | -10123 | -19691.5 | -14240 |
| UDF p=.020 r=.040 | 100.0% | 0 | 1.5 | 0 | 29075 | 0.04022 | 0.04647 | 0.04947 | -10123 | -19691 | -14240 |
| UDF p=.025 r=.050 | 0.0% | 157 | 2.5 | 0 | 18344 | 0.05018 | 0.05734 | 0.06096 | -9966 | -19690 | -14240 |
逐 case 输出
| case | variant | 选择 | watertight | boundary | components | nonmanifold | faces | proxy Chamfer | visible p95 | patch p95 | pitch | radius | OBJ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| gso_002_input4 | UDF p=.025 r=.035 | best | True | 0 | 1 | 0 | 29420 | 0.03602 | 0.04223 | 0.04624 | 0.025 | 0.035 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a40_local_implicit_completion/gso_002_input4/mesh_udf_pitch025_r035_trellis2_raw.obj |
| gso_002_input4 | UDF p=.025 r=.050 | candidate | False | 176 | 3 | 0 | 28616 | 0.05104 | 0.05800 | 0.06570 | 0.025 | 0.050 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a40_local_implicit_completion/gso_002_input4/mesh_udf_pitch025_r050_trellis2_raw.obj |
| gso_002_input4 | UDF p=.020 r=.040 | candidate | True | 0 | 1 | 0 | 46130 | 0.04103 | 0.04705 | 0.05314 | 0.020 | 0.040 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a40_local_implicit_completion/gso_002_input4/mesh_udf_pitch020_r040_trellis2_raw.obj |
| gso_008_input4 | UDF p=.025 r=.035 | best | True | 0 | 1 | 0 | 7212 | 0.03445 | 0.04076 | 0.04044 | 0.025 | 0.035 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a40_local_implicit_completion/gso_008_input4/mesh_udf_pitch025_r035_trellis2_raw.obj |
| gso_008_input4 | UDF p=.025 r=.050 | candidate | False | 138 | 2 | 0 | 8072 | 0.04932 | 0.05667 | 0.05623 | 0.025 | 0.050 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a40_local_implicit_completion/gso_008_input4/mesh_udf_pitch025_r050_trellis2_raw.obj |
| gso_008_input4 | UDF p=.020 r=.040 | candidate | True | 0 | 2 | 0 | 12020 | 0.03940 | 0.04589 | 0.04581 | 0.020 | 0.040 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a40_local_implicit_completion/gso_008_input4/mesh_udf_pitch020_r040_trellis2_raw.obj |
可视化结果


实验结论
- A40 给出正向机制证据:隐式表面比 A39 三角面缝合更适合处理 patch replacement 的拓扑闭合。
udf_pitch025_r035同时取得 boundary=0、nonmanifold=0、components=1 的平均结果,说明隐式重建能消除 A39 的非流形/碎片化病变。- 但它不是最终答案:UDF envelope 是无符号距离包络,会产生厚化、双层或平滑风险,visible surface p95 drift=0.04150。
udf_pitch025_r050半径更大反而出现 boundary=157,说明 UDF 参数不是单调更稳,后续要做 signed/local constraints。- 论文主线进一步清晰:不是 mesh 后处理拼接,而是 observable-state constrained implicit completion,并最终回到 TRELLIS.2 generation-time guidance。
下一步想法
- A41:signed/local SDF completion。用可见表面法线估计 inside/outside,只在 conflict tube 中更新隐式场,visible-locked 区域作为硬约束。
- 加入局部性指标:candidate surface 超出 conflict bbox/tube 的比例、visible-kept p95 drift、patch retention p95,防止 UDF 全局厚化。
- 把 A40 的 UDF envelope 与 A37 voxel closure 对比:二者都能水密,但 A40 有 replacement patch 约束,应验证是否更少破坏局部形状。
- 最终将 implicit completion signal 回投到 TRELLIS.2 sparse/SLat token,作为 generation-time observable-state loss。