-
Original U2Net Model updated to output 3 channel and loss function altered to L1Loss.
-
Training is done on Celeba-HQ dataset for ~150K iteration.
-
Final L1Loss on training dataset 0.060407.
-
Best model can be found here.
| Original Image | Masked Image | Model Result |
|---|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |