Closed ramge132 closed 1 year ago
In version 2 the effective backbone scale is more dynamic. In each intermediate unet latent, the distance between any feature vector and the mean is used to determine how much the backbone scale should be reduced. You can check the updated paper for a more precise explanation.
The recommended presets are not the same exactly for v1 vs v2. I'm waiting for the official code to receive an update before updating the code here.
At some point, version 2 became available, what's the difference?