<h2>KitchenSink V1.8 + 1.8L A SD 2.1 768+ Boosted Artistic General Model</h2>[[ this version is now archived and will not change. Mar 16th 2023 ]] [[ PLEAE USE VERSION 2 IN MY PROFILE. SFW/NSFW ]] Added LORA blend. GALLERY for GRIDS!! (Attachment limits sorry! Also civitai renames things I don't know why, on my pc it's named KitchenSinkVer1.8L etc. I suggest renaming to keep nice and clean after downloading) [ This model is a merge of all 2.1 based ~ 768+ models because I got sick of swapping between models and 512px is too low really. ] <ul><li>What the model differences are:1.8 = (the sloppy fun drunk) broad artistic vivid. Not scared to spill the wine while tipsy painting. Oh I made a Salvator Mundi? Cool.Example: The flower rose prompts. Vivid &amp; vibrant.1.8L ( for Lora blended in) = Less drunk with a sharpened crayon. Dinosaurs? ya gotchya, I won't make them as gross (Thanks dino lora's!) BUT it is much less vivid and punchy vibrant. </li><li>Base 2.1 !!! &gt; Don't forget the included YAML! (place in same folder as model. rename to same exact name as model! IE: KitchenSinkVer1.8 )[FYI You still need a yaml for 2.1 stuff - I don't know why some people say you don't anymore. For those reading it's less the auto11 not needing it and more that it is starting to give errors when merging/pruning/etc if you do not include your custom YAML file. -&gt; On this model please use my included sink yaml merci!] </li><li>Resolutions 768+ (768x768, 768x1024, 1024x768 ++ etc) </li><li>No clip nonsense, no vae nonsense, bog standard SD setup, simple to complex prompting as you wish, negatives not really required (fyi: the more negatives you enter without being excessive the better SD 2.1 preforms by the way!).IE: type cartoon if you want cartoon, realistic for real etc. call to it!! It has everything built in!! A few "realistic" keywords ? start with " masterpiece photo of... " and toss in "3d render, cartoon" into the negatives. =) </li><li>Sample gallery is random raw internet prompts on a fresh install - first try, no redo's, no backsies, sequentially numbered!, non cherry picked images. ( Fair is fair common! haha)</li><li>LOW Step count needed. Fantastic results in the 20-25 step range!</li><li>XYZ cat grid included to show 17,22,30 steps on all samplers (Standard CFG 7) "a cat lounging in the grass, basking in the morning sunshine" Note: some samplers do not work on (my) old gpu hence the few trash outputs. Apologies for those bits as it will probably work for you.)</li></ul>--&gt;&gt;[ SFW/NSFW - ??? totally untested (to the level some people want). As it's a total merge, anything is possible so don't be mad if something slides in - just add a negative word prompt to compensate. See merging info below. ] ** The model "seems" and feels generally safe for work. Just..yeah! unknown! =) <h2> WHAT IT LOVES AND HATES </h2>It loves &gt;??&gt; Flowers certainly so far! (Bloody brilliant!) It also seems to understands a wider artistic range of broad subjects and prompts. only needing light negative prompting is a bonus as well. It also puts an interesting sort of "oil paint realistic edge" ?! as it's own artistic style to things.It's good at &gt;++&gt; generally everything really ~ but more artistically than realistic from all the mixing. I randomized prompts from every site to make the gallery here and i find the results better in most cases when VS other general non purposely tuned model. (it seems to listen better)It's bad at &gt;--&gt; For one it lost the face/hands of the realistic models being diluted and has trouble with them in general now. And Secondly, It takes a few tweaks here and there. Meaning you might need to bump the CFG or steps depending on your purpose. Bit of fiddling is required depending on your purpose. Think ..tuning a guitar! see next point.--&gt;&gt; Not a bad point but really good to know: It's also a bit of a handful as larger models generally are, meaning sometimes you just need to run the same prompt a few times as it's got an eclectic mix of seeds deviating from the standard. I found that you cannot rely on a single result at times as it just hit a "bad" seed for your scenario. It's a hugely packed model that has an enormous seed resource pool built in! Making it a handful... completing the irritating loop haha. See the mixing below)Copyright &gt;&gt;&gt; I personally don't give 2 hoots what you do. Know that this has every model from every human that made one in 768+ rez until now mixed in. If it had a yaml it's in here. So, like, you probably gotta kinda sorta maybe follow their rules as it was originally their models. =) ** To be clear of what is in this model merge. All working non biased overcooked models in 2.1, with working yaml file, in 768+ resolution, from aug '22 to feb '23 are merged in. (Approx 700GB of models total) - which is why the pruned versions cannot be lowered past the 5GB size. I'm sorry it's as small as I can crush it down without clipping data. The version 1 description below has the long tedious formula I use. Post your comments and quibbles below.Please enjoy responsibly!BarQ! &gt; Hit the buymeacoffee link under my profile icon to help with the next R&amp;D project!&gt; Visit my artstation link to view custom built offline AI creations as well!

KitchenSinkVer1.8

******** HOW THE MERGING WAS DONE ********Understanding Merging (and why it takes long)Deck of cards merging analogyIn merging models you "shuffle them" together BUT keep the same number of cards. Take 2 decks of cards of different colors say Red and Black and shuffle your decks in real life with an attempt at 50/50 bias between them. BUT you also must end up with only 52 cards in your final deck after shuffling and you cannot look at the face values just the black/red colors.The results with anything but a pristine deck is going to be a bias towards one color or the other. poker time. Flip your cards. Do you have more jacks? two's? aces? nobody knows until you look and see (try the model) only to play poker with all ones...nice...a mayo deck. pffff.So after mixing you try models with the grid test xyz system across various prompts to expose it's weaknesses (you'll see in the added model screenshots)This is what I did. Mix some cards &gt; test the "new" 52 card decks &gt; using xyz + discarding the junk output models (Rats I shuffled all 1,2,3's when I was trying to make a high royal flush deck or something) junk that one. check the other 1/2 mix which turned out better? neither? Remix.It's a very tedious process!*********Base = Base SD2.1 768x768 modelbase mix was 50/50 weighted sum FormulaExample:Models A,B,C,DTake A+B - find best mix = ABTake C+D - find best mix = CDTake base2.1 + AB - find best mix = bABTake Base2.1 + CD - find best mix = bCDthen merge bAB + bCD etc etcmerge all using same method in month one until left with a base 1.0once complete, as you "lost cards" in the shuffle no matter what &gt; take all the models in this batched mix and change to ABC "add difference" type card shuffle merging. This takes KitchenSink 1.0 base at a rate of 0.2 and re-feeds in "lost" cards back into the core breaking the 52 card deck limit.Also making the model 28GB in size (yes gigs) when you are done re-adding all the missing bits. Loading it into the GPU memory is/was a nightmare heh.[The incremental versions, 1.1, 1.2, 1.3 etc are the month(s) rollup of the models from SD's rough start in august '22 until now -&gt; feb 23 2023 = 1.8 release for 8th month. Each month I would re-do the ABCD with only those models found in that period between themselves, then roll it into the previous kitchensink model. Incrementing upwards slowly having everything rolled into a ready to release model.][IE: all of this Marches models if any will be combined and merged into sink 1.8 to make 1.9 when the time comes. well feb 24th to march 23rd would be the next increment slice. haha]

Files with SHA256: 78dad762020f9624f2c7050e44668ca7e21b4e09f1c39e49cd267293b69f5b30