实验定位 方向定位正结果 / 方法尚未完成
A62 不声称已经完成最终方法，而是把 A9/A10/A11/A61 的证据合并成一个 budget placement audit：继续调后验 mesh selector 没有意义，单纯 token-count sampler band 也不够；真正有论文深度的下一步应该是 sampler-time observable-state budget。
最关键的三条证据是：A10 把 selected/control 的 token 偏差从 141.0 压到 1.0；A11 controlled full generation 仍然 ΔChamfer=0.0005456、ΔF@5=-0.002267、水密率=0.0%；A61 最好的后验 soft MV selector 仍比 A53 差 ΔChamfer=0.0004337。
实验设计（Image-2 风格）
模块设计（Image-2 风格）
Placement Audit
A62 把 budget 放置位置拆成四层：mesh 后验选择、sparse sampler token band、full generation token band、下一步 observable-state sampler budget。每层都只回答自己能证明的问题，不把稳定性结果包装成质量提升。
Observable State
下一步的状态不应只是 token count，而应包含 sparse support 几何、VGGT depth/silhouette 残差、visible/conflict/unobserved 局部标签，以及 decoder topology preview。控制动作可以是 soft correction、project-to-baseline 或 defer。
实验结果（表格）
Budget 放置位置对比
| placement | controls | positive evidence | limitation evidence | status |
|---|---|---|---|---|
| post_hoc_mesh_soft_mv_score | final candidate choice after meshes already exist | uses rendered multi-view depth agreement without GT | best no-GT selector a61_min_mean_mvd has ΔChamfer=1.6606540646481316e-05 while A53 has -0.0004171042648691843 | insufficient |
| sampler_time_token_support_band | per-step sparse token count relative to baseline trajectory | reset_count=2, selected abs token delta 141.0 -> controlled 1.0 | token-count band does not check observed depth, silhouette, or decoder topology | necessary_but_insufficient |
| full_generation_token_band | same support band carried through shape/texture/mesh decoding | seed1 empty selected sparse output is rescued back to baseline-like controlled mesh | controlled mean ΔChamfer=0.0005456264514101389, ΔF@5=-0.0022672011337224185, watertight=0.0 | stability_only |
| sampler_time_observable_state_budget | per-step token/support drift plus multi-view depth/silhouette consistency | addresses failures that post-hoc MV score and token-only band miss | not implemented yet; A62 only establishes necessity and design constraints | recommended_next |
失败模式覆盖
| failure mode | evidence | post-hoc MV | token support band | observable sampler budget |
|---|---|---|---|---|
| empty_sparse_collapse | A11 seed1 selected tokens=0 and selected generation failed; support-band reset_count=1 restored controlled tokens=26 | too_late | detects | should_detect |
| geometry_drift_without_token_violation | A11 seed4 selected tokens 1209->1210 with reset_count=0, but ΔChamfer=0.0010912529028202639 and ΔF@5=-0.004534402267444837 | weak | misses | target_case |
| topology_not_solved_by_sparse_budget | A11 baseline watertight=0.0/0.0; controlled watertight=0.0/0.0 | does_not_repair | misses | needs_decoder_topology_term |
| view_agreement_not_equal_global_3d_quality | A61 best soft-MV selector is worse than A53 by ΔChamfer=0.00043371080551566564 and ΔF@5=-0.0016466868275383075 | misses | not_applicable | must_be_soft_and_local |
A11 seed-level full generation 证据
| seed | baseline/selected/controlled tokens | resets | selected status | controlled status | selected ΔChamfer | controlled ΔChamfer | selected ΔF@5 | controlled ΔF@5 | controlled Δnonmanifold |
|---|---|---|---|---|---|---|---|---|---|
| 1 | 26 / 0 / 26 | 1 | failed | ok | - | 0.0000000 | - | 0.000000 | 0 |
| 4 | 1209 / 1210 / 1210 | 0 | ok | ok | 0.0010913 | 0.0010913 | -0.004534 | -0.004534 | 273 |
下一步方法约束
| constraint | reason | required signal |
|---|---|---|
| per_step_not_posthoc | mesh-level scoring cannot change a bad denoising trajectory after topology and surface have already formed | support token count, support bbox/centroid/IoU, and visible-depth residual computed during sparse sampling |
| observable_not_token_only | A11 seed4 has nearly unchanged token count but worse Chamfer/F@5, so count band misses geometry drift | VGGT depth/silhouette support projected into sparse support or SLat-visible regions |
| soft_local_not_hard_global | A60/A61 show global MV hard or soft scores can reject good edits or accept bad geometry | per-view/per-region residual with confidence weights and defer zones |
| decoder_aware | A11 controlled sparse stability still has watertight_rate=0, so sparse support alone does not solve topology | preview decoder topology budget or topology-carrier prior coupled to sampler states |
可视化结果
实验结论
- A62 的正结果不是“质量已经提升”，而是明确了贡献位置：预算必须放在 sampler trajectory 里，而不是 mesh 生成后的 selector 里。
- A10 证明 sampler-time support control 是可执行的：它能提前干预 sparse collapse，把 token 偏差大幅压小。
- A11 证明 token-only support band 不足：即使 sparse token 被稳定住，完整 mesh 的 Chamfer/F@5/topology 仍可能变差或不水密。
- A61 证明后验 MV soft score 不足：mean/worst/std MV depth 只能做诊断，不能当最终 selector。
- 因此下一步论文方法应定义为：Observable-State Guided Sparse Diffusion，把 VGGT 多视角 depth/silhouette 支持、局部可见锁和 decoder topology preview 一起变成 per-step 约束。
下一步想法
- A63：做最小 sampler-time observable budget prototype。在 A10 的 per-step hook 里加入 support bbox/centroid drift 和 VGGT visible-depth residual 的 proxy，而不是只看 token count。
- 先在 gso000 seed1/seed4 上复用 A11 full generation protocol，检查是否能同时保住 sparse 稳定性和 Chamfer/F@5。
- 把 hard project-to-baseline 改成 soft residual damping：局部超预算时降低 VGGT residual alpha，而不是整步回退。
- 增加 decoder preview topology term：在进入 shape SLat 前预测 boundary/nonmanifold/face budget，避免 sparse 稳定但 mesh 仍非水密。