# TRELLIS2 Hard-Part Crop Probe

- Image: `/data/250010098/physxanything_stage1_experiments/runs/omnipart_to_copart_partlevel_eval_20260411/input_images/00244988f3424ae0a7fb6f2381272361.png`
- Part id: `4`
- Overlap cleanup policy: `split_midpoint`
- Target bounds: `[[-0.25396825396825395, -0.23809523809523808, -0.49206349206349204], [0.23809523809523808, 0.2222222222222222, -0.03674603174603174]]`

## Criterion

- We want the hard part to succeed on the first decode, so the primary signal is `success` under a single crop config.
- Among successful configs, we compare `mesh/target bbox IoU` after the same soft-anisotropic fitting rule.
- We also track `effective_canvas_scale = object_fill_ratio * sqrt(mask_pixels / crop_area)` as a proxy for how visually dominant the target part becomes inside the TRELLIS input canvas.

## Results

| config | success | bg | pad | fill | alpha | eff_canvas_scale | faces | IoU | aniso |
| --- | --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| baseline_dim | no | dim_context | 0.15 | 0.72 | 96 | 0.2999 | n/a | n/a | n/a |
| transparent_zoom | yes | transparent | 0.10 | 0.88 | 0 | 0.4365 | 80216 | 0.4685 | 17.2201 |
| white_zoom | yes | white | 0.10 | 0.88 | 255 | 0.4365 | 744746 | 0.6200 | 7.3976 |
| dim_lowalpha_zoom | yes | dim_context | 0.10 | 0.88 | 24 | 0.4365 | 49148 | 0.3502 | 182.9583 |
| transparent_wide | yes | transparent | 0.22 | 0.86 | 0 | 0.2898 | 107310 | 0.4938 | 12.4934 |

## Best Config

- `white_zoom` with IoU `0.6200`, background `white`, padding `0.10`, fill `0.88`, alpha `255`.