diff options
| -rw-r--r-- | README.md | 1 |
1 files changed, 1 insertions, 0 deletions
@@ -53,6 +53,7 @@ Ideas of mine that did not work unfortunately: * Efficientnet was apparently a terrible choice of an encoder - A ConvNext module heavily copied from lucidrains [x-unet](https://github.com/lucidrains/x-unet) + was incredibly much better at encoding the images to a better representation. * Use VQVAE to create pre-train a good latent representation - Tests with various compressions did not show any performance increase compared to training directly e2e, more like decrease to be honest |