ModelsLab/Mangled Merge XL - V3.0

It is my pleasure to introduce version 3.0, the next iteration of the Mangled Merge XL series. I've spent some time looking into the DARE/TIES model merging method and holy moly what a difference! This model is a continuation of Mangled Merge XL V2.0 with an additional 600 loras merged in with the DARE/TIES method making this model clock in at a total of 1800 loras! Once I get a better grasp on a reliable smoothing method, I am thinking about starting from the beginning and remerging with a more improved (and automated) workflow.The model excels in just about all styles, output is much improved from V2.0. Furry animals are fixed, 1girl and booru tags work, and the model is really good at pixel art as well. Whatever style you are looking for, it helps to put it at the very beginning of your prompt. The first 3 example images show what I mean in their respective prompts. CFG scale is good between 1.5 and 6.5.The main method for merging loras was done in ComfyUI with the help of the ComfyUI-DareMerge nodes by 54rt1n. I added attn_only normalization to the lora+base model and used that as an attention gradient along with a magnitude masker between the base and model B to perform an Advanced/DARE merge at 0% drop with 2 iterations and rescale on. Then I injected gaussian noise into the base model with ratio: .98, mean: .05, and std: .01 and merged that into the first merge with a Block/DARE merger node and using the same magnitude masker with a .9 drop rate, rescale off, 1 iteration, only to the input block. I also replaced the clip from version 2.0 with the clip from the base and didn't change it during the merging process which makes for much better prompt adherence. This method allowed me to merge about 180 models before I started to see anything wonky happening.The smoothing process was done by hand every 150 merges. I haven't found a method that I can use every time for automation just yet but I hope to have one by the next version. For the most part the smoothing method consists of taking the starting model (0), then the model 1/3 in (50) the model 2/3 in (100) and the last merge (150), then doing a simple merge on each one with 10% of the base to smooth, then I merged the 4 together with the Attention/DARE node in various combinations and weights until I found the most balance in test runs. The order in which the models were merged and the weights were different with every 150 merges. Align Your Steps was used on the scheduler during tests and for all example images.Enjoy!