Replies: 1 comment 1 reply
-
@Corshi wow! sounds great. I wish we can see classes and Identifiers. So that we can pick and choose to merge only the trained data, like style, person, etc. can you please tell, in which of "25 U-Net blocks" dreambooth trained data like person or style exists? |
Beta Was this translation helpful? Give feedback.
1 reply
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
The example of implementation is in there https://github.com/bbc-mc/sdweb-merge-block-weighted-gui
It just gives more control over what is being merged. It still requires a trail and error, but idea of finding "block" that will transfer specific things like "texture" or "anatomy", combined with the [re-basin] seems like a very nice combination.
Just to clarify. The extension in the link separates model into 25. 12 starting blocks, middle block and 12 ending blocks.
And with the manual setting, you can recreate the effect of smootchstep or other blending algorithms to blend merges even better.
OR as it seems, skip the blocks that do accumulate the distortions and corruptions (if found in testing etc.).
Last thing. Thanks for this awesome gui!!
Beta Was this translation helpful? Give feedback.
All reactions