OmniPart x TRELLIS2 · experiment detail
2026-04-17 20:24:36 UTC

OmniPart x TRELLIS2 Formal Soft-Fit Selected Probe

把正式 soft_anisotropic per-part pipeline 扩展到三类代表性样本,并用联合判据评估:mean mesh/target bbox IoU、condition overlap risk、decode stability。这个页面更接近我们真正要盯的东西:哪些结果是几何贴框成功,哪些只是被重叠 bbox 或恢复重试掩盖了问题。

OmniPartTRELLIS2soft-anisotropicformal-probeassembly-diagnostics
2026-04-17 20:24:36 UTCTimestamp
3Assets
activeStatus
OmniPart x TRELLIS2 Formal Soft-Fit Selected Probe cover image
Assets
Interactive Asset

000ec112ae7f4a8a93f847ccfd4031be - mechanical sample

机械形样本。两个 parts 都在第一次 crop/decode 就成功,mean_IoU=0.6512,overlap_risk=0.0299,decode_stability=1.0。这里的主要瓶颈不是条件框冲突,而是第二个 part 仍有明显几何弱化,说明 crop 语义和 part 内部细节仍在掉。

0.6512mean IoU
0.0299overlap risk
1.0000decode stability
000ec112ae7f4a8a93f847ccfd4031be - mechanical sample Overview
Overview
000ec112ae7f4a8a93f847ccfd4031be - mechanical sample Part 00 crop
Part 00 crop
000ec112ae7f4a8a93f847ccfd4031be - mechanical sample Part 00 soft-fit preview
Part 00 soft-fit preview
000ec112ae7f4a8a93f847ccfd4031be - mechanical sample Part 01 crop
Part 01 crop
000ec112ae7f4a8a93f847ccfd4031be - mechanical sample Part 01 soft-fit preview
Part 01 soft-fit preview
Mechanical Formal Soft-Fit Preview
软各向异性拟合公式仍为 u=mean(log(e_b/e_s))+(log(e_b/e_s)-mean(log(e_b/e_s)))/(1+lambda),lambda=0.25。这个样本说明当 boxes 不重叠时,当前正式链条已经能稳定闭环,但局部 part 质量仍未完全站稳。
Interactive Asset

0023687e90394c3e97ab19b0160cafb3 - flower thin-structure sample

花卉细结构样本。两个 parts 一次成功,mean_IoU=0.7860,overlap_risk=0,decode_stability=1.0,是目前三例里最干净的正式成功案例。它说明当前 baseline 并不是一概不行;当 OmniPart boxes 干净、crop 语义相对纯净时,TRELLIS2 的 per-part 生成是能稳定站住的。

0.7860mean IoU
0.0000overlap risk
1.0000decode stability
0023687e90394c3e97ab19b0160cafb3 - flower thin-structure sample Overview
Overview
0023687e90394c3e97ab19b0160cafb3 - flower thin-structure sample Part 00 crop
Part 00 crop
0023687e90394c3e97ab19b0160cafb3 - flower thin-structure sample Part 00 soft-fit preview
Part 00 soft-fit preview
0023687e90394c3e97ab19b0160cafb3 - flower thin-structure sample Part 01 crop
Part 01 crop
0023687e90394c3e97ab19b0160cafb3 - flower thin-structure sample Part 01 soft-fit preview
Part 01 soft-fit preview
Flower Formal Soft-Fit Preview
这是当前最值得反复观察的一例,因为它同时包含薄结构和较高 bbox agreement,却没有靠重试或严重重叠撑住结果。
Interactive Asset

00244988f3424ae0a7fb6f2381272361 - box-like nested-layout sample

盒状家具样本。表面上 mean_IoU=0.7705,看起来比机械样本还好,但这是一个很危险的假象:target bbox 自身 overlap_risk=0.5469,而且第二个 part 需要第二次 crop 才解码成功,decode_stability=0.6667。也就是说,这例真正暴露的是条件布局歧义和解码脆弱性,而不是单纯的贴框问题。

0.7705mean IoU
0.5469overlap risk
0.6667decode stability
00244988f3424ae0a7fb6f2381272361 - box-like nested-layout sample Overview
Overview
00244988f3424ae0a7fb6f2381272361 - box-like nested-layout sample Part 00 crop
Part 00 crop
00244988f3424ae0a7fb6f2381272361 - box-like nested-layout sample Part 00 soft-fit preview
Part 00 soft-fit preview
00244988f3424ae0a7fb6f2381272361 - box-like nested-layout sample Part 01 crop
Part 01 crop
00244988f3424ae0a7fb6f2381272361 - box-like nested-layout sample Part 01 soft-fit preview
Part 01 soft-fit preview
Box-Like Formal Soft-Fit Preview
这是目前最重要的反例。它说明仅看 fitted bbox IoU 会误判质量,因为条件框本身就可能在装配空间里强嵌套。