Mangled Merge XL

Magic:Magic is the first of a two part experiment in making specialized models. To make Magic, I took Mangled Merge XL v4.0 and merged 500 2D based style loras (2800 total now) into it using MBW Dare/TIES with 50% into input block 8 and output block 0, and 100% into output block 1. Ties were set to count and gradient was the method used. This first version was focused on 2D while the next model will be focused on realism. I plan to merge the two together afterwards for Mangled Merge XL v5.0.This model does well within a CFG range of 2-7. For more realism, go for the lower ranges, while the higher ranges work better for 2D, machines, or more complicated prompts. You will also see better outputs when putting styles first in your prompt.Enjoy!V4:This update has 2300 loras merged in total. That's 500 more than V3. I didn't smooth this last one quite as much which has led to a little bit more bias towards photo/hyper realism. I decided not to smooth for styles in this one as I really enjoyed seeing the difference and added detail in some of the outputs.The model seems a little more sensitive now but also more powerful and random from my early tests. Unless you're trying anime styling, it's best to steer clear from booru tags like 1girl, or score_whatever. Also, the model seems to give the most attention to the beginning tags so you can use them for styles, although it is more difficult to get illustrations in this one. It seems like it's more versatile in other ways now however. Some CFG scales have changed for certain subjects and I highly recommend using tools like AYS, FreeU, and PAG to really get the most out of your outputs.Enjoy!V3: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!V2:I am happy to introduce Mangled Merge XL version 2.0 to the Civitai community. This new version includes 1200 loras that have been carefully merged in to "train" the model. That's 380 more than the previous Dark version! Mangled Merge XL 2.0 also includes Hillside, Clarity, Colorful, Realistic Stock Photo, Pixelwave, Realism from Hades, Pilgrim 2D, and Crystal Clear.Usage:Mangled Merge XL does well with a CFG range between 2 and 7. There are some tokens that look much better with a low CFG score, but you can also go between 5 and 7 for the model to follow your prompt more closely if it is more complicated.The model tends to stick more to realism but can do paintings and illustrations, it's just a little harder to get them in it's current state. I want to balance it out more but I thought 1200 was a nice number to release the model with. There was/is no goal to make this model more one way or another as I am mostly working towards prompt adherence and aesthetics.Mangled Merge XL 2.0 also likes full body type long shots. It does close up portraits but if you don't specify, it will tend to output closer to full body or further away pictures. Luckily, eyes and faces come out great for this model with distance shots. Hands are decent but not perfect. But you can always touch up with some inpainting.If you are going to use strong weights, I suggest lowering the CFG or else outputs may look a little fried. Also "fur" and "1girl" can give weird outputs. Fur needs a low CFG and 1girl can potentially give unreliable outputs. It's better to use "a woman" or something along those lines. The model is prompt sensitive so it will output what is in your prompt and for the most part it doesn't output what's not in your prompt. So if you don't specify a background, it is likely to give you a white or otherwise plain background.Disclaimer:If you are trying to reproduce the example images I have provided, I am using the AI Diffusion extension for Krita which has some generation options that aren't available to add to the Civitai prompt information. Civitai only allows whole numbers for in the CFG scale inputs, but some of these models are 6.5 or 5.2 or something like that, so you may need to dial the CFG scale in to get the exact output. Civitai also doesn't have an option for the sampler I used which is Turbo/Lightning Merge - DPM++ SDE which requires the 8-steps lora. All were produced with 12 steps rather than 8 however because I thought it gave better outputs. All images were also upscaled with an additional refining process.I think that covers about everything. Have Fun!V1.1 DarkQuick update with an additional 80 loras to help give little more strength to tokens like "chiaroscuro lighting, and low key" for more dramatic lighting in the output imagesV1:Mangled Merge XL is the product of an experiment to see if it was possible to "train" a base model by merging loras into it and what method to use that prevents the model from breaking. After merging approximately 760 loras and 3 trained models (Cinevision, Juggernaut, and HaveAll), the model is still stable for more merges and gives beautiful outputs.Mangled Merge XL tends to do well with a CFG score between 3 and 8. The goal was to make it a good general use model made with a broad range of loras that I thought looked pretty cool. With that said, there is nothing explicit in here as far as adult material. There is nudity but that's about it. If you would like to have a more NSFW focus, feel free to merge it with your favorite NSFW models.The model also doesn't do very well with strongly weighted tokens or crazy resolution sizes. Sticking to the native SDXL resolutions will give nice results however.Note: All sample images were hit with a refined upscaling process.Have fun! Any feedback is welcome.Merging process:There isn't anything too complicated or scientific that went into the process. It was basically trial and error to see what the best results were and how to keep everything stable enough to continue merging without the model breaking. I didn't go crazy with block weights or anything like that, but feel free to build upon the following process to produce even better models!1. Merge 10 loras into an SDXL 1.0 base model using the Merge Lycoris tool in Kohya_ss. 50% weight is usually fine here but feel free to adjust.2. Take the 10 lora model and merge it back into the SDXL base model using the Merge Checkpoint feature in Auto1111 to smooth out anything that went awry during the merging process in step 1. A lora will hold it's style throughout the merging process and also can have strong weights so after merging 10 loras, your model might give broken outputs. This step will smooth out those crazy weights to keep the model from breaking. Nothing crazy is needed here. I used weighted sum and the weights were generally 10 Lora 65% / SDXL Base 35%, but sometimes after testing the outcome, I would move the weights for the 10 Lora model down depending on how crazy the model got during the 10 lora merging process.3. Make a copy of the smoothed model. Hold on to one and merge 10 new loras into the 2nd copy.4. Smooth out the new 20 lora model using the instructions from step 2 and then merge the 2nd smoothed model (60%) with the 1st smoothed model (40%) to make your first Amalgam.5. Copy the Amalgam and merge 10 loras into it. Continue the process in steps 1-4 until you get your 2nd Amalgam.6. Merge Amalgam 2 (60%) with Amalgam 1 (40%) to make your first Amalgamerge.7. Continue the process in steps 1 - 6 until you get your second Amalgamerge.8. Merge Amalgamerge 2 (60%) with Amalgamerge 1 (40%) to make Amalgamerge 39. Continue until you get another Amalgamerge and merge that with Amalgamerge 3.10. And so on...Every now and then, things can go a little more awry than expected. If that happened, I would merge with a select trained checkpoint in order to try and fix some tokens that were getting too wacky.