Sloppy Moments

Planting kernels... ...with Fizgig. Did some test runs with Fizgig Anima training implementation and merged a couple of resulting loras into some slop. Namely, assorted painting and assorted bara loras trained on anima-aesthetic. It's nothing you haven't seen before, truth be told, and then probably a little less. But it's quirky enough and deserves an upload. As to output - anima doesn't really do trad painting textures very well, so it's sort of has digital softbrush look to it. And you still need to use painting juice to get it. That aside, it's a generic bara lora. Seems consistent from a few tests. Another thing worth noting - Fizgig tags data with natural language. Does it make a difference? No clue. Enjoy. Update Replaced the lora with a more diluted version. It was a little too rough on the edges and a little below of what I consider pretty or interesting. New version's general output feels a little 'DIY dataset' styley, but '' tag is a lot more stable in return. Dizdoodz inspired lora A prolific homoerotic artist. Anima is very aware of his existence. I blended it, as usual, for a little more consistency, I guess. But it's still quite raw'ish. (tbf, realisticbeasthero compliments this a lot better, but I didn't want to make another variant of that) You can use '', if you want to - I didn't train on it, but it exists in Anima and enhances the style. A few backhanded comments on the art style itself. Original art is very refined, neat, clean and sports ridiculous dynamic poses and interactions. But when you start generating basic '1man, standing, flexing biceps' - it kinda looks like most average western-anime mix ever with some flat ink shading sprinkled on top, minus puffed up naughty parts. But hey, at least it's not a "50 shades of hair" waifu lora! Also, looking back, should have probably separated line art from colored art - output is very random. Oh well. Enjoy! A few prompts Over at realisticbeasthero style I uploaded a small set of colorful dynamic prompts, that I use to generate hunk abstractions. Keywords entirely arbitrary and superficial. They work on some models and not at all on other. But, since I'm uploading abstract images a lot, I thought I'd have them all in one place for anyone to try. Have fun. Japanese Caricature Style First of all, I started training on Anima-Aesthetic-v1.1 now. Don't see any drawbacks, and models end up stronger, I feel, from a couple of tries so far. This model was trained on over a hundred images of expressive "super deformed" Japanese caricatures for 10 epochs. It still very much produces general output and if anything is a good anime lora. For "super deformed" output use your general keyword (that shall not be named). Japanese caricatures are somewhat controversial lately, as such - a declaimer: This model is NOT intended to produce NSFW "super deformed" content. Do not produce it and do not upload your images. Also, I did remix it a little as usual, so it's not too rigid. Enjoy. 2 parts of slop + 1 part of kim hyung tae Another lora inspired by one of better unique oldtimer styles. Don't even know what medium his art was used in - never seen any of it. But HUGE thighed chicks burned into my cranium since a very long time ago. While original training data was predominantly female, 2 parts of slop were homoerotic. No idea if any female knowledge left. Didn't work with any less slop, unfortunately. Kim Hyung Tae happens to be a grand master of irl AI-hands. There're almost no images with clearly visible 5 fingers. Still, end result lora is flavorful in its own right. Enjoy. Merge it harder, make it better, post it faster, upvote stronger... Remember the oldie Tastefully Nude? Those were the days... Trying to make something out of nothing. Ah... Anyways, I uploaded a new male dataset with a terrible lora trained on it and got visited by 3 Ghosts of Merging. Long story short, I bring you a first pure lora merge (with nothing new added) since like forever. Inspired by a recent series of images made by mixing realistic, hero and beastman styles. Anima merging is quite annoying, but, in the end, I landed on something quite similar. Enjoy. By popular demand! "A stack of bad photography training"! Turns out render style lora from Anima Preview1 days is getting attention still! So, I remade it for Anima Base. The original thing contained deconstructron-photography, deconstructron-render art, toon style and a mysterious bit called 'maleabc' (all P1). This one contains deconstructron-render art (P3), toon style (P3) and a brand new 'maleabc' (Base)! What's so good about this? - Glam handpicked sportie guys! - Greater facial variance than normal. Little to no resemblance to training data, but faces are different per gen. - Default realism-leaning style and strong(ish) render style. Biggest con - broken gens are an eyesore on realistic output, and Anima sprinkles errors generously. I do, normally, try to make my loras work at weight 1, but don't forget you can still adjust the weight or mix this with something else, more stable perhaps. Anyways, enjoy! Sometimes a prince is just not enough! If you know where last one is from - you know where this one is from too :winkyface: Very straight forward. - There may or may not be a two letter keyword that enhances the style (even though, I think this should enhance all American comic book styles across the board through similarity and association). - Conflicting ("word salad") prompts are poor, because data had a lot of "weak signals" in form of auto-tagged comic panel pages. Prompts have to be focusing "strong concepts" for best result (like, you will find that certain sex poses are super consistent, while a random solo dynamic pose will be entirely broken). (for those unaware how Anima text encoder works: Anima uses AI text encoder that infers general "idea" of your prompt and only then looks for latent association; unlike, say, SD1.5 encode