- BiRefNet Portrait
- Open weights · commercially permissiveLicense: MIT
- U²-Net Human Segmentation
- Open weights · commercially permissiveLicense: Apache-2.0
birefnet-portrait vs u2net_human_seg
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 | birefnet-portrait | u2net_human_seg |
|---|---|---|
| MGE (Edge Quality) | 0.05950 ✓ | 0.08689 |
| MAE (Overall Accuracy) | 0.05032 ✓ | 0.09628 |
- MGE
- 0.14119✓
- MAE
- 0.13362✓
- MGE
- 0.18039
- MAE
- 0.26527
- MGE
- 0.06774✓
- MAE
- 0.22680✓
- MGE
- 0.06967
- MAE
- 0.22743
- MGE
- 0.04515✓
- MAE
- 0.00868✓
- MGE
- 0.09961
- MAE
- 0.03528
- MGE
- 0.01512✓
- MAE
- 0.00591✓
- MGE
- 0.04501
- MAE
- 0.02340
- MGE
- 0.06204✓
- MAE
- 0.03288✓
- MGE
- 0.08194
- MAE
- 0.06321
- MGE
- 0.05246✓
- MAE
- 0.02145✓
- MGE
- 0.06375
- MAE
- 0.03799
- MGE
- 0.04472✓
- MAE
- 0.01411✓
- MGE
- 0.05262
- MAE
- 0.06527
- MGE
- 0.16214✓
- MAE
- 0.10275✓
- MGE
- 0.19185
- MAE
- 0.24021
- MGE
- 0.02311✓
- MAE
- 0.00370✓
- MGE
- 0.08905
- MAE
- 0.11216
- MGE
- 0.02896✓
- MAE
- 0.00886✓
- MGE
- 0.05164
- MAE
- 0.02787
- MGE
- 0.08324✓
- MAE
- 0.04343✓
- MGE
- 0.09263
- MAE
- 0.05819
- MGE
- 0.05236✓
- MAE
- 0.01182✓
- MGE
- 0.10398
- MAE
- 0.03386
- MGE
- 0.01095✓
- MAE
- 0.00289✓
- MGE
- 0.03566
- MAE
- 0.01279
- MGE
- 0.06671✓
- MAE
- 0.02252✓
- MGE
- 0.07784
- MAE
- 0.02957
- MGE
- 0.15394✓
- MAE
- 0.04942✓
- MGE
- 0.16755
- MAE
- 0.20554
- MGE
- 0.01953✓
- MAE
- 0.00800✓
- MGE
- 0.05299
- MAE
- 0.03977
- MGE
- 0.01972✓
- MAE
- 0.01254✓
- MGE
- 0.04393
- MAE
- 0.03770
- MGE
- 0.13269✓
- MAE
- 0.22283✓
- MGE
- 0.17383
- MAE
- 0.37089
- MGE
- 0.03465✓
- MAE
- 0.00886✓
- MGE
- 0.07626
- MAE
- 0.05722
- MGE
- 0.07836✓
- MAE
- 0.32167✓
- MGE
- 0.08966
- MAE
- 0.32712
- MGE
- 0.00660✓
- MAE
- 0.00225✓
- MGE
- 0.02200
- MAE
- 0.01058
- MGE
- 0.01762✓
- MAE
- 0.00610✓
- MGE
- 0.04587
- MAE
- 0.02406
- MGE
- 0.09258✓
- MAE
- 0.02708✓
- MGE
- 0.12973
- MAE
- 0.23347
- MGE
- 0.02527✓
- MAE
- 0.01801✓
- MGE
- 0.04527
- MAE
- 0.04463
- MGE
- 0.01406✓
- MAE
- 0.00223✓
- MGE
- 0.04340
- MAE
- 0.00965
- MGE
- 0.00956✓
- MAE
- 0.00312✓
- MGE
- 0.03144
- MAE
- 0.01333
- MGE
- 0.01959✓
- MAE
- 0.00804✓
- MGE
- 0.04789
- MAE
- 0.03824
- MGE
- 0.03003✓
- MAE
- 0.00879✓
- MGE
- 0.06552
- MAE
- 0.03158
- MGE
- 0.04145✓
- MAE
- 0.01784✓
- MGE
- 0.09699
- MAE
- 0.11154
- MGE
- 0.15421
- MAE
- 0.21967
- MGE
- 0.15004✓
- MAE
- 0.13050✓
- MGE
- 0.22170✓
- MAE
- 0.10099✓
- MGE
- 0.26149
- MAE
- 0.19606
- MGE
- 0.03258✓
- MAE
- 0.00972✓
- MGE
- 0.05666
- MAE
- 0.04061
- MGE
- 0.04777✓
- MAE
- 0.01845✓
- MGE
- 0.07874
- MAE
- 0.08652
- MGE
- 0.01511✓
- MAE
- 0.00599✓
- MGE
- 0.03927
- MAE
- 0.03209
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).