OmniPart x TRELLIS2 · experiment detail
2026-04-17 20:45:38 UTC

OmniPart x TRELLIS2 Overlap-Cleanup Box Probe

专门拿最棘手的 box-like 样本做条件清洗实验。我们把原本重叠严重的 3D boxes 用 split_midpoint 规则拆开,再走同一条正式 TRELLIS2 per-part 生成链,检查 overlap risk、retry burden 和最终装配会不会一起改善。

OmniPartTRELLIS2overlap-cleanupbox-likefailure-analysis
2026-04-17 20:45:38 UTCTimestamp
1Assets
activeStatus
OmniPart x TRELLIS2 Overlap-Cleanup Box Probe cover image
Assets
Interactive Asset

00244988f3424ae0a7fb6f2381272361 - baseline vs split_midpoint cleanup

这是一次很有价值的负例对照。cleanup 把 overlap_risk 从 0.5469 降到了 0,但 retry_burden 仍是 1,decode_stability 仍是 0.6667。也就是说,条件歧义被清掉了,但最难的第二个 part 仍然需要 retry,说明真正的瓶颈已经更靠近 crop / decode 本身。

0.5469 -> 0.0000overlap risk
1 -> 1retry burden
0.7705 -> 0.7366mean IoU
00244988f3424ae0a7fb6f2381272361 - baseline vs split_midpoint cleanup Baseline overview
Baseline overview
00244988f3424ae0a7fb6f2381272361 - baseline vs split_midpoint cleanup Cleanup overview
Cleanup overview
00244988f3424ae0a7fb6f2381272361 - baseline vs split_midpoint cleanup Cleanup part 00 crop
Cleanup part 00 crop
00244988f3424ae0a7fb6f2381272361 - baseline vs split_midpoint cleanup Cleanup part 00 fit preview
Cleanup part 00 fit preview
00244988f3424ae0a7fb6f2381272361 - baseline vs split_midpoint cleanup Cleanup part 01 crop
Cleanup part 01 crop
00244988f3424ae0a7fb6f2381272361 - baseline vs split_midpoint cleanup Cleanup part 01 fit preview
Cleanup part 01 fit preview
Cleanup Result Preview
这页默认展示 cleanup 后的交互式 GLB,同时把 baseline 和 cleanup 的 overview 并排放进图集。重点不是看它有没有变漂亮,而是看条件清洗究竟解决了哪一层问题、又没解决哪一层问题。