# OmniPart + TRELLIS2 Assembly Diagnostics

- Run: `/data/250010098/physxanything_stage1_experiments/runs/omnipart_trellis2_overlapcleanup_box_probe_20260417/00244988f3424ae0a7fb6f2381272361`
- Success parts: `2 / 2`
- OmniPart boxes: `4`

## Part Stats

| idx | part_id | crop | target volume | mask px | faces | mesh/target IoU | fill ratio |
| --- | --- | --- | ---: | ---: | ---: | ---: | ---: |
| 0 | 1 | dim_context | 0.187577 | 12033 | 1206698 | 0.8089 | 0.9483 |
| 1 | 4 | dim_context | 0.103132 | 3240 | 789844 | 0.6642 | 0.9373 |

## Pairwise Target BBox Relations

| part_i | part_j | bbox IoU | intersection volume | gap |
| --- | --- | ---: | ---: | ---: |
| 1 | 4 | 0.0000 | 0.000000 | 0.010000 |

## Notes

- These diagnostics are AABB-level sanity checks, not surface-intersection tests.
- High target bbox IoU means assembly conditions overlap before any mesh generation.
- Low mesh-to-target bbox IoU or very anisotropic extent ratios can indicate poor bbox fit or part hallucination.