Update README.md
Browse files
README.md
CHANGED
|
@@ -1,9 +1,15 @@
|
|
| 1 |
-
Semantic VAE
|
| 2 |
|
| 3 |
Trained from DINOv2 in 12 hours as follows:
|
|
|
|
| 4 |
Frozen DINOv2 -> LayerNorm -> 32 channels
|
|
|
|
| 5 |
Frozen DINOv2 with patch embed unfrozen -> LayerNorm -> 32 channels
|
|
|
|
| 6 |
64 channel bottleneck with Sigreg loss (1e-3 weight) from LeJEPA
|
|
|
|
| 7 |
VA-VAE convolution decoder
|
|
|
|
| 8 |
Loss is all layers of DINOv2 and DINOv3 equally weighted MSE on input vs reconstruction.
|
|
|
|
| 9 |
There are no pixel losses/LPIPS/GAN used, this is DINO space loss only.
|
|
|
|
| 1 |
+
#Semantic VAE
|
| 2 |
|
| 3 |
Trained from DINOv2 in 12 hours as follows:
|
| 4 |
+
|
| 5 |
Frozen DINOv2 -> LayerNorm -> 32 channels
|
| 6 |
+
|
| 7 |
Frozen DINOv2 with patch embed unfrozen -> LayerNorm -> 32 channels
|
| 8 |
+
|
| 9 |
64 channel bottleneck with Sigreg loss (1e-3 weight) from LeJEPA
|
| 10 |
+
|
| 11 |
VA-VAE convolution decoder
|
| 12 |
+
|
| 13 |
Loss is all layers of DINOv2 and DINOv3 equally weighted MSE on input vs reconstruction.
|
| 14 |
+
|
| 15 |
There are no pixel losses/LPIPS/GAN used, this is DINO space loss only.
|