well9472 commited on
Commit
4de4887
·
verified ·
1 Parent(s): 1c9bf60

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +7 -1
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.