Adaptation of the merging method described in the paper - Git Re-Basin: Merging Models modulo Permutation Symmetries (https://arxiv.org/abs/2209.04836) for Stable Diffusion
Calculates average of newL - oldL for each permutation. Then takes the weighted sum of the two permutations using the difference in average newL - oldL. Seems to be a little better than the original method with the downside of being slightly slower.
I can't test this right now, so please let me know if it works better.
Calculates average of newL - oldL for each permutation. Then takes the weighted sum of the two permutations using the difference in average newL - oldL. Seems to be a little better than the original method with the downside of being slightly slower.
I can't test this right now, so please let me know if it works better.