unnamed girl from tuki. <a target="_blank" rel="ugc" href="https://www.youtube.com/watch?v=AOfaWrBwo7I">Official Music Video Clip</a> Animation by <a target="_blank" rel="ugc" href="https://x.com/240_mado">十八番茶 (240_mado)</a> Also, this girl in these clips: <a target="_blank" rel="ugc" href="https://www.youtube.com/watch?v=j8iGq4TO5Fg">ギルティ</a>and <a target="_blank" rel="ugc" href="https://www.youtube.com/watch?v=N0zZKq3RCQo">最低界隈</a>(But my lora doesn't contain a single frame from there, at least not yet. I'm glad to see how the singer's singing and the artist's animation have improved.) All images have a prompt you can look at and choose what you are interested in!CHECK OUT THE RECOMMENDED CFG FOR ILLUSTRIOUS!Also if you're confused by the different version numbers, etc. You can look at the quality of the loras from left to right.General:Trigger words: t1u5Also only on v0.92 (16 dim) NAI-XL*: s8t, flat color explanations can be found below in the details of the version.Recommend words:For default girl: t1u5, 1girl, mole under mouth, colored sclera, pink eyes, symbol-shaped pupils, black hair, colored inner hairFor default girl + hair and jewellery: t1u5, 1girl, mole under mouth, colored sclera, pink eyes, symbol-shaped pupils, long hair, black hair, hime cut, sidelocks, earrings, bracelet, hair stick, hair ornament, single hair bun, colored inner hairfor full girl attributes: t1u5, 1girl, mole under mouth, colored sclera, pink eyes, symbol-shaped pupils, long hair, black hair, blue kimono, sleeveless, earrings, bracelet, red nails, hair stick, hime cut, sidelocks, hair ornament, colored inner hair, yellow obi, red sleeves, detached sleeves, hakama short skirt, pale skin, red leg ribbon, platform footwearMay use like trigger/useful words:flat color for the original character style (don't forget to try raise the weight of the lora.)gradient caption when a character has noticeable traces of color change on the clothing below, you can put that word in negative prompt, I would consider that colors an artifact in most cases.gradient background caption as it should be when the background is gradient. But you can also use it as a negative word to reduce artifacts.BUT both words do not generate artifacts by themselves and can be used in a positive prompt as well.you might also try: japanese clothes, bare shoulders, getaGeneral possible problems/Non-obvious points:The amount of hair stick will always be different this is a feature of sdxl, the mole on the face will change position sometimes. You can try to get a mole on the left side of the character's body on the collarbone, it's in the dataset and caption, but it will be difficult and unstable I think. hakama short skirt, that's how that white part of the clothes is captioned, it varies quite a bit (as I think in the clip this part also changes so) yellow obi changes its size so much because in the clip the size also changes depending on how far away the character is. That black yellow thing, in the form of a big belt or I don't know what, is also there, but it only appears as an artifact and doesn't have a special caption. blue kimono caption exactly like that, not sure what it might affect, but you should know and like pink eyes, symbol-shaped pupils a character's eyes color cannot be called just pink, but I decided to go with this, if you are trying to achieve a special pink color, or a different symbol in pupils, it is recommended to change the weight of the word and add an additional prompt. The blue kimono sometimes becomes a dress underneath because the original clip does the same thing.Many small details are captioned, such as lighting, gradient, faceless and etc.Also the “fan girl” character is available inside, with the right facial features selected it can be pretty stable.The 32 dim versions and v0.92* is worse in models, but they contain many shots where there is no main character. black skin might work differently here and produce a different result. My tests didn't show that it helps to highlight the other characters well there, but I guess you can get them with a better chance from them, the stability of the style will also be higher, but its are also a bit more artifacted.The differences between pony xl and Illustrious/NAI XL:pony has all its standard pros, but this time I didn't find much pros outside of different models and loras. Style, anatomy, details, I think everything is better on NAI, you can try and decide for yourself. Pony is still quite good and it's not much worse. Illustrious/NAI XLRecommended weight: 0.6-1.2 (0.9) for more original style 1/1.1 Recommend cfg scale: 2.5-4 (3)at least for <a target="_blank" rel="ugc" href="https://civarchive.com/models/926443?modelVersionId=1061268">this noobai xl model.</a> v0.97 (16 dim) NAI-XLThis version is the best option for noobai similar models, because there is no style dataset and the character is much more flexible to adapt to the style, as well as because of the lack of unnecessary data. But the character is still poorly adapted to styles, so it is recommended to increase the weight of styles, as well as writing combinations of them, perhaps flat color in the negative will help to reduce the impact of the original style a little bit.v0.96 (32 dim) NAI-XLThis version has a more stable style but is also more artifacted, you can notice some gradient at the bottom. As mentioned above, the version has more images not related to the character. Together, the version has more detail and perhaps better adapts the original style. But overall, it doesn't offer any particular advantages.v0.92 (16 dim) NAI-XL*as you can tell even by the number, this version is special. Regardless, I liked her work, so I kept it.Features: the version has the most images in the dataset, including the style.s8t trigger word where minimum character images (if and works very weakly) You can try to use t1u5, s8t to get something interesting or make the style more stable.flat color used as trig words almost everywhere to try to extract and concretize the style. It's not particularly successful.But you can try to give the word more attention here, and also put it in the negative prompt if you're trying to get rid of the original style.Because of all this, in theory the model should be more flexible than the others, in practice I haven't noticed it. The model is quite artifactual and has not always optimal signatures. You can consider it experimental.Trained on <a target="_blank" rel="ugc" href="https://civarchive.com/models/833294?modelVersionId=1022833">NoobAI-XL Epsilon v1</a>Pony XLv0.95 feature: as you can see here are different epoch of the same version, so it is different training duration, so epoch 8 you can use by default, but for more precise details and original style (e.g. character's skin turning blue) you can use epoch 12.v0.95 epoch 8 16 dim PDXLRecommended weight: 0.7-1.2 (0.8-0.9) for more original style 1more flexible, all the traits of the character are internalized, but no more.v0.95 epoch 12 16 dim PDXLRecommended weight: 0.7-1.2 (0.9-1)Even features like blue skin, blue lines are starting to be internalized, maybe the model is a little more artifacted because of this.v0.93 (32 dim) PDXLRecommended weight: 0.7-1.2 (1) for more original style 1.1Possible problems/Non-obvious points:Again 32dim everything is the same as the others, but there is a special feature here, the way of training is different from all the others. It's more standard for me. Not sure what it affects exactly, but it is noticeable, the model has brighter colors, especially when increasing weight, and also reacts differently to increasing weight. Also at the same weight, the style in the model is not as noticeable, so as written above, the weight is higher than the others. Not sure when exactly you should use it.Trained on Pony diffusion v6Recommended to use adetailer, hiresfix.And so, as I said before, this model was made with similar ideas to <a target="_blank" rel="ugc" href="https://civarchive.com/models/964218/ashlin-character-alienmonday-2-outfits-xenomorph-pony-illustriousnai-xl">this one</a>. By itself, almost at the same time.But unlike the last version, the ideas were simpler and maybe realized even better. As you can see in some of the images, there is more than just the character. The style here is more unique, I guess that's why it's more flexible. But it's still not a style model. I thought about making one, but I don't have the time.Also because I stretched the creation of the model for more than 3 months, someone made the character a bit earlier. <a target="_blank" rel="ugc" href="https://civarchive.com/models/965701/tuki-hyururirapappa">This one.</a> I'll be honest, I like my versions better. But you can compare for yourself. I throw here a big comparison of all versions of pdxl and you can download and see difference, version compressed to 50mb because the original weighed 217mb.Let me tell you right away THIS IS A BIASED COMPARISON. Despite the fact that there is a trigger word of that lora, but some words in prompt are used according to my caption. For example pink eyes, whereas the other author has purple eyes, etc. Also, the relative weight of lora is probably different, that's why it shows the result at different weights lora.

OC girl from tuki.『ひゅるりらぱっぱ』by 十八番茶 mado [Illustrious/NAI, Pony XL] HYURURIRAPAPPA 最低界隈 ギルティ

32 dim trained on Adam8bit than others on CAME.

tukiGirlPDXL32v0.93tx1.safetensors

v0.93 (32 dim) PDXL

Files with SHA256: 4795b67979237efdc5fe6450436811cb759f517dfd4d015653fbc1b413bf8cdf