- withoutBG Open Weights
- Open weights · commercially permissiveLicense: Apache-2.0
- BiRefNet HRSOD
- Open weights · commercially permissiveLicense: MIT
withoutBG Open-Weight Model vs birefnet-hrsod
Quantitative comparison: 34 images scored against ground-truth mattes. Lower is better. Use the view controls on each row to switch between transparency, chroma key, and alpha matte on all outputs.
| Metric | withoutBG Open-Weight Model | birefnet-hrsod |
|---|---|---|
| MGE (Edge Quality) | 0.03911 ✓ | 0.05566 |
| MAE (Overall Accuracy) | 0.01241 ✓ | 0.03030 |
Disclosure: withoutBG includes BiRefNet
withoutBG Open Weights and Pro Model API are routed pipelines. A classifier sends fine strands, soft detail and transparency to withoutBG's own matting model, and hard opaque objects, flat scenes and vehicles to BiRefNet (MIT license), which is part of the pipeline. On this page, 11 of 34 images went to BiRefNet, so on those images this page partly compares BiRefNet with BiRefNet. The table splits the scores by route: the first group is what withoutBG's own matting model handled.
| Images | Metric | withoutBG Open-Weight Model | birefnet-hrsod |
|---|---|---|---|
| Routed to withoutBG matting (23) | MAE | 0.00893 ✓ | 0.02503 |
| MGE (edges) | 0.02915 ✓ | 0.03850 | |
| Routed to BiRefNet (11) | MAE | 0.01969 ✓ | 0.04132 |
| MGE (edges) | 0.05993 ✓ | 0.09152 |
Per-image averages, lower is better.
- MGE
- 0.04103✓
- MAE
- 0.00678✓
- MGE
- 0.10666
- MAE
- 0.03756
- MGE
- 0.05151✓
- MAE
- 0.02894✓
- MGE
- 0.08378
- MAE
- 0.12069
- MGE
- 0.05343
- MAE
- 0.00944✓
- MGE
- 0.05336✓
- MAE
- 0.00989
- MGE
- 0.01719✓
- MAE
- 0.00444✓
- MGE
- 0.02760
- MAE
- 0.01410
- MGE
- 0.02392✓
- MAE
- 0.00652✓
- MGE
- 0.03619
- MAE
- 0.01169
- MGE
- 0.04106✓
- MAE
- 0.01152✓
- MGE
- 0.04905
- MAE
- 0.01791
- MGE
- 0.03247
- MAE
- 0.01071
- MGE
- 0.01962✓
- MAE
- 0.00544✓
- MGE
- 0.07023✓
- MAE
- 0.01136✓
- MGE
- 0.09296
- MAE
- 0.01683
- MGE
- 0.02835✓
- MAE
- 0.00423✓
- MGE
- 0.02913
- MAE
- 0.00554
- MGE
- 0.02826✓
- MAE
- 0.00782✓
- MGE
- 0.03945
- MAE
- 0.01697
- MGE
- 0.04685✓
- MAE
- 0.00925✓
- MGE
- 0.05330
- MAE
- 0.01138
- MGE
- 0.06053✓
- MAE
- 0.01170✓
- MGE
- 0.06630
- MAE
- 0.01412
- MGE
- 0.02119
- MAE
- 0.00506✓
- MGE
- 0.01989✓
- MAE
- 0.00551
- MGE
- 0.04161✓
- MAE
- 0.01017✓
- MGE
- 0.04531
- MAE
- 0.01049
- MGE
- 0.05709✓
- MAE
- 0.01017✓
- MGE
- 0.14535
- MAE
- 0.05671
- MGE
- 0.02022✓
- MAE
- 0.00706✓
- MGE
- 0.03191
- MAE
- 0.01630
- MGE
- 0.01419✓
- MAE
- 0.00564✓
- MGE
- 0.02268
- MAE
- 0.01404
- MGE
- 0.08276✓
- MAE
- 0.02340✓
- MGE
- 0.13722
- MAE
- 0.11387
- MGE
- 0.04719
- MAE
- 0.01842
- MGE
- 0.02933✓
- MAE
- 0.00535✓
- MGE
- 0.02469✓
- MAE
- 0.00504✓
- MGE
- 0.04154
- MAE
- 0.00969
- MGE
- 0.01337✓
- MAE
- 0.00710
- MGE
- 0.01345
- MAE
- 0.00483✓
- MGE
- 0.02154✓
- MAE
- 0.00550✓
- MGE
- 0.03087
- MAE
- 0.01259
- MGE
- 0.03721✓
- MAE
- 0.00546✓
- MGE
- 0.05176
- MAE
- 0.00892
- MGE
- 0.03279
- MAE
- 0.02348✓
- MGE
- 0.02781✓
- MAE
- 0.02580
- MGE
- 0.02246✓
- MAE
- 0.00368✓
- MGE
- 0.02812
- MAE
- 0.00556
- MGE
- 0.01142✓
- MAE
- 0.00267✓
- MGE
- 0.01795
- MAE
- 0.00599
- MGE
- 0.02087✓
- MAE
- 0.00589✓
- MGE
- 0.02646
- MAE
- 0.00951
- MGE
- 0.02626✓
- MAE
- 0.00684✓
- MGE
- 0.09121
- MAE
- 0.23094
- MGE
- 0.03987✓
- MAE
- 0.01721✓
- MGE
- 0.07902
- MAE
- 0.06336
- MGE
- 0.13077
- MAE
- 0.09352
- MGE
- 0.12241✓
- MAE
- 0.04154✓
- MGE
- 0.09317✓
- MAE
- 0.01618✓
- MGE
- 0.13557
- MAE
- 0.02569
- MGE
- 0.02545✓
- MAE
- 0.00694✓
- MGE
- 0.04528
- MAE
- 0.02559
- MGE
- 0.03566✓
- MAE
- 0.01269✓
- MGE
- 0.06522
- MAE
- 0.03561
- MGE
- 0.01505✓
- MAE
- 0.00707✓
- MGE
- 0.02651
- MAE
- 0.02019
Methodology
- Alpha mattes are grayscale masks encoding per-pixel foreground probability. They capture fine edge detail (hair, fur, semi-transparent regions) that binary masks discard.
- Metrics are computed by comparing predicted mattes against a ground-truth reference: MGE measures edge sharpness and MAE measures pixel-level accuracy. Both are lower-is-better.
- Scored images come from Bench v0, where every image has a ground-truth matte. The benchmark page has the full methodology and per-image scores for every model (CSV / JSON).