<h3 id="small-update">Small update</h3>Uploaded 10 epoch version.Training data somewhat more pronounced in this one.But it's a bit of a tometo tomato situation.You can find the trigger tags over there ------------&gt;Tags are not mandatory, you can use overarching general tags I mentioned before, like 'sketch', 'lineart', 'realistic', 'painting', 'greyscale'.- Do take note, that tag 'realistic' is often THE difference between anime output and western output.- Dragons and unicorns should work too, to a degree. Animals lora data was included.- 'devchar' for anime'er output.- Yes, I know tags are all over the place, but I can't change them post training.<h3 id="enjoy.">Enjoy.</h3><hr /><h3 id="bad-news-everyone!">Bad news everyone!</h3>...there will be no finetune............<h3 id="instead-there's-a-combined-lora!">Instead, there's a combined lora!</h3>It's fresh out of the oven, so the details are still muddy. We'll test it together as we go.But here are a few bullet points that will help you decide if you care:- It's most of everything minus 'photography'. Tried to train a merged lora including photography a bunch of times, but it's a giant waste of time. Photography never looks good and affects everything else negatively. This version is entirely devoid of photos.- 'render art' is ends up being mostly meh, but I kept it in, and it's the most realistic looking thing you can get from this.- Contains some of the stuff from my other Anima loras, like animals or ink style. I'll update trigger tag list a bit later, after I test if there's even any point to trigger tags.- Has all the usual tags like 'traditional art', 'pencil art', 'bara art' and such. BUT, as you may or may not know, diffusion models like to reuse latent space. Meaning tiles don't always land where you intend them to (without Text Encoder training). To evoke actual pencil art, you have to do 'lineart', 'monochrome', 'sketch' and it will generate what it was finetuned on under 'pencil art'.- As such Anima has some very powerful tags that suck up most of art medium identity, like 'sketch', 'lineart', 'realistic', 'painting'.- For stronger flat anime looks try 'toon style' or 'devchar' custom tags.- Trained on Anima - Base. 5 epochs of 5000 images.Will update this as I go. Maybe even cook it a little more if needed.<h3 id="good-luck-and-have-fun!">Good luck and have fun!</h3><hr /><h3 id="preview3-versions">Preview3 versions</h3>Everything is pretty much the same as before.The only difference being render-art. Added a raw version this time, instead of a "beauty remix".Mix them together.Add Toon Style, Takeshi Matsu Style or something else entirely to top it all off.Generate sexy hunks.Enjoy and good luck!<hr /><h3 id="a-small-snack...">A small snack...</h3>...while I tinker with other models.Classical concept art style.Anima - Preview3Other models are still Preview1. But highres output in Preview3 is very juicy, so I will be updating some.(unet only trained - no trigger needed, but you have to watch out with describing background and 'solo' tags)<hr /><h2 id="big-slop-dump">Big slop dump</h2>I didn't get to figuring out Anima finetuning yet, so I trained individual loras on parts of Sigmatron.I've mentioned before - assorted datasets train very similarly on Anima. As such style differences are quite subtle.But data, being unique sets of 200-500 images, made output vary quite a bit per lora.<h3 id="what-is-this-for">What is this for?</h3>Remixing mostly - with other loras, artist tags or whatever else. Loras do work on their own and produce pretty pictures, but there's artstyle randomness to it - you'll need to anchor it a little if you want consistency.As per usual, datasets were only male. Anima can probably generalize to some degree, didn't try.Also, I'm not reuploading stuff that's already out, namely:- chargen art (My Sloppy Husbando something something)- toon style- animals- render artYou can find the above in my profile, if you need those.<h3 id="7-loras-total">7 Loras total</h3>They're named after styles, but don't take it too seriously. Mediums don't quite show unless you bake it hard or prompt engineer for it. Maybe, if I trained it longer, medium identity would show more, but image to image style delta might also just push and pull and never land on a common ground.("style delta" - that's a term from 21st century)Photography is a standout, because it's very week on its own. Just use it as an addon for more realism to other stuff (it's contained in render art lora, for example).Pencil art is retrained with more epochs from original upload.Grid previews are all the same seed - a couple missed.<h3 id="do-you-need-all-of-them">Do you need all of them?</h3>Probably not. They're picture slot machine loras. You know, for fun.So yeah, that's it for now.Maybe I'll do a finetune later, when technology is here.Until then.

Deconstructron

deconstructron-art-10.safetensors

deconstructron-base-10e

deconstructron-base-5e

photography-P3

yaoi-art-P3

bara-art-P3

traditional-art-P3

render-art-P3

pencil-art-P3

illustration-art-P3

digital-art-P3

concept-art-P3

photography-P1

bara-art-P1

yaoi-art-P1

traditional-art-P1

pencil-art-P1

illustration-art-P1

digital-art-P1

Files with SHA256: abd49c9d6e53cdea89a4855342db56618031a8a16c1198bcdef2a403a46adb6c