Follow me on X/Twitter for exclusive content: https://x.com/Citron_Legacy 😁
Check out all the Art styles I’ve made: https://civarchive.com/user/CitronLegacy/models?tag=style
This Lora is trained on 859 images so it probably has all the characters. Here are the characters I know work:
Meidri - Meidri, 1girl, bangs, large breasts, brown hair, long sleeves, dress, frills, wings, puffy sleeves, orange hair, apron, maid, maid headdress, juliet sleeves, monster girl, feathered wings, puffy long sleeves, white wings, flipped hair, harpy,
Crimvael - Crimvael, solo, long hair, open mouth, blue eyes, blonde hair, dress, bare shoulders, wings, shorts, white dress, strapless, halo, black shorts, bike shorts, strapless dress, angel wings, angel, shorts under dress
Hyena girl - long hair, blush, large breasts, navel, animal ears, cleavage, bare shoulders, blue hair, swimsuit, bikini, hand on hip, short shorts, black bikini, denim, bikini top only, furry, blue shorts, denim shorts, furry female,
Blue Demon - 1girl, solo, long hair, bow, navel, yellow eyes, red hair, wings, horns, pointy ears, midriff, bowtie, detached collar, colored skin, demon girl, colored sclera, demon wings, oni, blue skin, black sclera
Dark Fairy - 1girl, solo, long hair, breasts, smug, blonde hair, large breasts, dress, cleavage, bare shoulders, jewelry, earrings, green hair, wings, pointy ears, black dress, tattoo, piercing, fairy wings,
Angel Hyena merge - 1girl, solo, long hair, breasts, large breasts, animal ears, cleavage, blue hair, swimsuit, brown eyes, upper body, bikini, wings, black bikini, parody, furry, cigarette, furry female, white wings, smoking, body fur, animal nose, snout, smug, smile,
Witch - 1girl, solo, smile, looking at viewer, long hair, breasts, looking at viewer, blush, smile, bangs, blue eyes, blonde hair, large breasts, simple background, gloves, hat, white background, dress, cleavage, bare shoulders, very long hair, pink hair, braid, multicolored hair, one eye closed, elbow gloves, white gloves, blunt bangs, large breasts, aqua eyes, single braid, witch hat, blue dress, blue headwear, witch,
Slime girl - slime girl, 1girl, solo, long hair, large breasts, closed mouth, green eyes, green hair, indoors, colored skin, monster girl, bookshelf, green skin,
Mushroom witch - 1girl, solo, long hair, large breasts, hat, cleavage, brown eyes, collarbone, grey hair, black dress, hair over one eye, tree, witch hat, pale skin, nature, forest, one eye covered
Description
FAQ
Comments (6)
The Kiff Style???
Seems rather overtrained and badly tagged. Enforces long hair or stomach bulge everywhere. Gave it a like anyways.
Thanks for the feedback. This is the exact same dataset I used for the Illustrious version, but I guess that means I should use less repeats/steps/epochs for Anima.
This is the 10th epoch, so maybe the solution would be to use the 7th epoch? Did you think it was acceptable at a lower weight like 0.8?
As for tagging, I simply used the tags from WD1.4 and prefixed character names, but that was a long time ago. Lately I've been tagging with SmilingWolf_wd-eva02-large-tagger-v3 so maybe I should retag all of the images?
Definitely interested in your thoughts!
@CitronLegacy unfortunately I cannot call myself an expert. This is a long going holy war to tag images or not tag images, use tag dropout or not, use short tags or long tags, use activation word or not etc. In my limited experience I had to use wd swin as base and then cleanup stuff manually. Right now I am building dataset for style and maybe try a thing or two.
Unfortunately I did not use your illustrious lora and did not plan to, so cannot really compare. I was more interested in looking how anima loras are performing at all. Especially considering no official training code or even diffusers implementation is in sight
Sorry in advanced for the wall of text!
@Volnovik In this field its hard to say what qualifies as "expert" because things are changing and is there even a "correct" way to make art? 🤷♂️ Personally I want to learn more so if someone does something different I want to know why. For example, I never learned how people make detailer/enhancer loras.
When tagging datasets, I always start with automatic tags and then manually prune or add tags for control/details. For this style I had no particular goals so I only added character names for a few characters. My best models are trained on datasets that have a lot of custom or enhanced art, but this lora is not one of them. 🤣
Usually when I switch to a new base model I like to make 50 or so loras before I really understand the right balance of training settings. So sadly that means my first few loras are sacrificed for science.
As for training code, I guess its not official but I'm using this https://github.com/kohya-ss/sd-scripts/pull/2260 .
Its too bad about the lack of diffusers. I trained almost all my Pony and Illustrious loras on diffusers so it was a bit of an adjustment to not use diffusers for Anima.
Anyway expert or not your works are well known. I see your NoobAI-XL Detailers fairly often. When you made your NoobAI-XL Detailers what techniques did you use? How many images?How much manual tagging did you do? Do you edit/enhance your images before training? I won't pry into trade secrets but if you have any tips I'd definitely like to hear them.
@CitronLegacy author of anima has his own training sw diffusion-pipe. It looks like it is the only "official" implementation that we will get.
About detailer - all guidance is spread thin in comments and description. I trained a slider lora for details (0.5 version). In ostris, text only. It was an ass to make it run tbh, other than that be sure to put equal amount of tags into positive and negative.
Then I generated a dataset of around 400 1024x1024 images using the model itself. I do not remember tags, but it was like mimialism, flat color, underdetailed etc. Around 10 per concept covering wideshots, close-ups, scenery, landscape, 1boy, 1girl etc. and trained a lora using Prodigy. Tagging was WD and cleanup after. It was rather easy, since there we really few objects on the scene. Then I skimmed through undertrained versions to have the most aesthetically pleasing result at negative weight. That was painful, but rewarding. Then I played around with combining it with 0.5 version and merged 2 loras in supermerger with desired weights. That's it, nothing fancy. You can consider that a poor man's SPO 🤣 since I basically use weights of model to cancel undesired results. That's why my model does not affect style (if it is not flat color minimalism obviously). Sometimes I open a beer and scroll through comments about my lora having bad dataset with slop hair, greasy skin and realistic imagery to have a laugh. Btw you can "see original" imagery if you use only my lora with weights like -10 or -15, I do not remember.