My first LoRA because I like Grea and didn’t see one for SDXL/Pony yet. <a target="_blank" rel="ugc" href="https://civarchive.com/models/282341?modelVersionId=629577">4th Tail Model v0.4.5</a> helps a lot for complex tail concepts like holding tail and more.<h3 id="tags-you-should-consider-adding-in-positivenegative:-cyqmpx4yc">Tags you should consider adding in positive/negative:</h3>See sidebar first for default outfit and bikini tags; Grea's Light outfit from GBF can be found <a target="_blank" rel="ugc" href="https://civarchive.com/models/758201">here as it's own LoRA</a> (Compatible with other characters) P: For v3 it's pretty much a requirement to add <code>dragon tail</code> in the prompt (see "About this version" for info)P: HAIRSTYLE: <code>short hair</code> or <code>long hair, hair over breasts</code> also <code>hair censor</code>helps. And: <code>ahoge</code> if you want it; looks cute :&gt;"Stabilizer" tags like <code>dragon wings, dragon horns, pointy ears, center frills</code> (and any others) not needed but use them if they are missing (model dependent); dragon-tags in negative might not work, untested; center frills if they disappear (usually when doing color change) N: <code>shoulder straps</code> when prompting high waist skirt (depends on model but they appear often from my testing because of the high-waist skirt)N: <code>breast pocket</code> to prevent said concept form appearing on the shirt/nightgownN: <code>lace-trimmed legwear</code> or (P)<code>plain thighhighs</code>if you encounter frilled/lace-trimmed thigh highs although from my testing they weren't nearly as bad as in v2 (i actually forgot to tag them in v3 so they got trained in, my bad)P/N: <code>high-waist</code> remove it or add it based on your preferences. Controls skirt somewhatP/N: <code>lace-trimmed bikini</code> (weak) for the blue-ish "fins"/lace-trim from the anime bikiniP: <code>anime screencap</code> could trigger the anime style depending on the model used. (untested on pony, but works very well with the <a target="_blank" rel="ugc" href="https://civarchive.com/models/826063">Illustrious</a> / <a target="_blank" rel="ugc" href="https://civarchive.com/models/826063">Tofu version</a> of this LoRA)Note: o-ring bikini might not work correctly in cowboy/full body shots because not enough data (merges with known but closeup/upper body should work <h3 id="flexibility-(colorsconceptsstyles):-nuonkteal">Flexibility (Colors/Concepts/Styles):</h3>Overall the LoRA should be pretty flexible. As in changing hair or eye color or style shouldn't be a problem although sometimes you need to add more weight or put the default color in negative when a color change isnt working. The high-waist skirt from default outfit seems to need extra work with color change. The wings seem to be trickier ("blue wings" for example made them completely blue but sometimes spilled over to the tail and "black wings" just kept the membranes at default color). Color change example (Black themed Grea): Here is an example of positive and negative tags i used for the black clothed/scaled grea (this was just a quick test to see if it's possible, i don't know if this is too many tags or not but it worked including the side ponytail hairstyle): P: <code>black horns, black wings, black dragon tail, black scales, black hair, dark purple eyes, black shirt, black high-waist, black pleated skirt, black plaid skirt</code> N: <code>red hair, purple hair, red plaid skirt, red high-waist, red pleated skirt, blue skirt</code>For more info and technical info, check "About this version".To support me without tipping Buzz, use the "Add post" button above the gallery (instead of "Post images" from the "Create" dropdown) if you want to publish images made with my LoRA. Each post earns me 50 Buzz. You can do this for other creators too.Side notes:<ul><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/288584">AutismMix</a> sometimes gets the anime style quite accurate</li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/448716?">BoleroMix</a> has great style and the author has a ton of clothing/concept lora worth checking out</li></ul>Update note 4th Oct.: There will be a 3.1 version at some point where i fix some of the tagging to cause less bleeding like o-ring bikini, but there aren't many images of the bikini so it might not even do anything. If I decide to work on a 3.1 then maybe I'll tag <code>tail under clothes/skirt</code> and see if it affects how other models treat Grea's tail and if it brings more control (see winter black capelet showcase image using t-ponynai3, tail is not under skirt and suggests back cutout/merge). If this test doesn't work out I'll just leave it as it is.

Grea (Multiple Outfits) | Manaria Friends / Granblue Fantasy / Shingeki no Bahamut | (Pony)

I learned more! most likely.Cleaned the dataset and tags a LOT for a better result (mainly stability across concepts/outfits) and it is now easier/more efficient to prompt an outfit through less tags needed to do so. Important: see the model description for tags to include or exclude, other than those, do that as usual with objects you (don't) want in your image Multiple outfits available! (See sidebar info trigger words)Main: Default (default white shirt) and Bikini (beach episode, o-ring might have issues, if so just remove any occurance of it in your prompt) Extra: Nightgown, Winter coat with capelet Both <code>short hair</code> and <code>long hair</code> Grea should be pretty stable from possibly all views, have fun!Long hair style is much more stable now as well as hair over breasts and hair censor when the model itself doesn't have enough knowledge of it.Why dragon tail is a requirement but other tags like dragon wings/horns isn't:<code>dragon tail</code> in prompt is a MUST if you want it in the image because I tagged it in the dataset since without that tag the tail would often merge with clothes or just not appear at all even with it prompted, which would lead to several more issues such as the likeness being gone. This is only a requirement for dragon tail and does not apply to the horns, wings or hair/eye color. To change those just put whatever you don't want in negative and to change eye color simply prompting "blue eyes" or similar.Hair style/color change may require higher weight. Side note:As already hinted, the <code>grea</code> activation trigger includes the wings, horns, eye color and pointy ears (and probably the hair color, which might be absorbed by short hair or long hair too, which is why hair style/color change might require higher weight). It should also include the breast size (medium-big) because I couldn't be bothered to readd the tag even programmatically and Pony is good enough to make them bigger/smaller when prompted anyway. Fun Fact: the size varies throughout the anime surprisingly often anyway.I never tried outfit tags but decided against them because I thought they would make the outfits less flexible.Technical stuff:The model was trained with base PonyXL v6 on <a target="_blank" rel="ugc" href="https://github.com/bmaltais/kohya_ss">kohya_ss</a> Network Dim rank 32; alpha 16 (no dreambooth, don't have the vram lol; 3060 12gb). The LoRA was resized to rank 8.If you are interested in the settings I used (if you know about training loras, please tell me if there's anything to be improved!): <a target="_blank" rel="ugc" href="https://files.catbox.moe/m1klvs.json">https://files.catbox.moe/m1klvs.json</a>Quick note to that: I used prodigy and 6100 steps. the reason it didn't overtrain is probably because of scale weight norms is set to 5.note if it matters: I use kohya_ss with pytorch 2.4.1+cu124 and xformers==0.0.28.post1 <s>xformers==0.0.28.dev893+cu124 (found </s><a target="_blank" rel="ugc" href="https://github.com/facebookresearch/xformers/actions/runs/10559887009"><s>here</s></a><s>; install with --no-deps to prevent from torch 2.4.0 getting installed and then edit the METADATA file of xformers dist-info folder in lib/site-packages to require torch &gt;=2.4.0 instead of torch ==2.4.0 kudos to </s><a target="_blank" rel="ugc" href="https://github.com/facebookresearch/xformers/issues/1096#issuecomment-2339039663"><s>this guy</s></a><s>)</s> irrelevant now that xformers 0.0.28.post1 is out 12th Sep. 2024 - edited 27th Sep 2024

Files with SHA256: 346c91ef36cd3605eafcfba009be94aed5586c55d1f30228d5fa2fa808e0c771