Illustrious:
Testing Illustrious XL. I was uploading funny capybara images and suddenly they got "Blocked" and "Failed to upload image". After a while I became able to upload images again. I removed the prompts from the two images that were blocked. Anyway, when using Euler Ancestral CFG++ for Illustrious XL, I recommend using a different sampler for HiresFix. Same dataset as before, BTW. I haven't cleaned anything up.
Illustrious XLを試し中です。面白いカピバラの画像をうｐしたら、いきなり「ブロック」と「失敗」のエラーになりました。しばらく待ったらうｐできるようになりました。ブロックされた二つの画像のプロンプトを消しました。とにかく、Illustrious XLにEuler Ancestral CFG++を使ってるなら、HiresFixの時に別のサンプラーを使うほうがおすすめです。ちなみに、前と同じデータセットです。まだ何も掃除してません。
BTW this doesn't show up in thumbs if you search for "Illustrious" despite it being in the title, but it does if you search "Illustrious micro" or limit date rate.
ちなみに、タイトルに「Illustrious」書かれてるのに、そう検索したらサーチに出ません。「Illustrious Micro」か日付範囲で検索したら出ますが。
Original Text:
I liked this 1girl from a while back and thought, why not try making a LoRA of her? So I did. This is essentially a one image dataset. I upscaled and then cut it up into many parts. It works better than I expected, so I decided to share it. V5 has a more varied dataset, so it should be able to do more things, and also adhere to other styles better. V5 can also be used for the outfit only. Note: I slightly changed some prompts for better results.
I recently started experimenting with SDXL AutismMix Pony. Results seem mixed. Sometimes great, sometimes bad. "Full body" seems to give worse results than SD1.5. I'll slowly try redoing my old LoRAs. If I get a good result I'll upload it. This one was trained in Pony. It also works in AutismMix Pony.
この1girlさんが前から好きで、LoRAにしてみようと思いました。画像1枚で作ったデータセットです。その画像を大きくして何枚に分解しました。結果は思ったより良さそうだから、うｐしようと決めました。V5のデータセットの種類がかなり増えたので、できることが増えて、絵のスタイルも影響されやすくなったはずです。V5は衣装だけにでも使えます。お知らせ：もっと良い結果になるためにタグを少し変更しました。
最近SDXLオーティズムミックス・ポニーを使って実験始めました。結果がたまにすごくいい、たまにあまりよくない。全身の絵なら、SD1.5より結果が良くないみたいです。今までのLoRAを少しずつやり直してみます。良い結果が出来たらうｐします。今回はポニーで作りました。オーティズムミックス・ポニーにも使えます。
Recommended weight （おすすめウエイト）:
v1.0: 0.8 - 1
v5.0: 0.5 - 1
v5.0 (outfit only/衣装だけ): 0.5 - 0.6
I recommend using LoRA Block Weight with MIDD, so that the art style will be more like the checkpoint model you choose.
LoRAブロックウエイトの「MIDD」おすすめです。そうしたら、絵のスタイルはもっと選んだチェックポイントのようになります。
Example (例）: <lora:MiniskirtGirl_v5:1:lbw=MIDD>
Recommended prompts（おすすめタグ）:
Try these when parts are not displayed properly. (いろいろ正しく表示されてない時にこれを使ってみてください)
・v1.0: earrings, jewelry, collared shirt, white blouse, black hair, long hair, black necktie, sleeveless, black corset with gold buttons, underbust, high-waist skirt, red-pink skirt, pleated skirt, miniskirt, lowleg panties, micro panties, string panties, black panties, side-tie panties, black fishnet thighhighs, platform heels, platform footwear, toeless footwear, strappy heels,
・v5.0: black hair, black eyes, long hair, braid, asymmetrical bangs, sleeveless shirt, high-waist skirt, frilly skirt, very short skirt, pink skirt, black corset, gold buttons, lowleg, string panties, micro panties, side-tie panties, fishnet thighhighs, back-seamed legwear, long necktie, black necktie, platform heels, strappy heels, toeless footwear, black footwear, platforms,
・v5.0 (outfit only/衣装だけ）: sleeveless shirt, high-waist skirt, frilly skirt, very short skirt, pink skirt, black corset, gold buttons, lowleg, string panties, micro panties, side-tie panties, fishnet thighhighs, back-seamed legwear, long necktie, black necktie, platform heels, strappy heels, toeless footwear, black footwear, platforms,
・PonyXL v1: black hair, black eyes, long hair, braid, asymmetrical bangs, sleeveless shirt, high-waist skirt, frilly skirt, very short skirt, pink skirt, black corset, gold buttons, lowleg, string panties, micro panties, side-tie panties, fishnet thighhighs, back-seamed legwear, long necktie, black necktie, platform heels, strappy heels, toeless footwear, black footwear, platforms,
・PonyXL v1 (outfit only/衣装だけ）Untested but presumably same as v5.0 まだ試してないけどv5.0と同じのはずです。
・IllustriousXL v1: mskgrl, french braid, swept bangs, black hair, (long hair:1.5), black eyes, brown eyes, medium breasts, (sleeveless:1.5), white blouse, collared shirt, (long black necktie:1), (underbust:1), black corset, (gold buttons:1.4), black nails, high-waist skirt, reddish pink skirt, (very short skirt, pleated skirt:1.5), skindentation, string panties, lowleg panties, side-tie panties, (wide hips:1.3), (fishnet thighhighs:1.3), black footwear, platform heels, (strappy heels, toes:1.5),
Recommended negative prompts（おすすめネガティヴタグ）:
To help draw parts properly. (いろんなパーツを正しく表示するため)
・cleavage, collarbone, white thighhighs, very long hair, medium hair, garters, garter belt, red eyes, tied hair, ponytail, pink underwear, sleeves, red footwear, red platforms,
Personal preference examples (好みによって使えるタグの例):
・bare legs, bare feet, covered nipples, see-through, large breasts, thicc thighs, skindentation, wide hips, side-swept bangs, short hair, medium hair, braid, side braid, pale blue eyes, grey eyes, holding handbag, earrings,
See gallery for better examples. (ギャラリーでもっと良い例が見れます)
Problems (問題点):
Patterns and minor details may be off.
パターンと細かい詳細がたまにずれてます。
v1.0: Sometimes draws her calves shorter as if to make up for the platform heel height.
プラットフォームヒールの高さの代わりに脹脛が短くなる場合がみたいです。
v1.0: Sometimes draws the colors wrong.
色がたまに違うように表示されます。
v1.0: Hi-res Fix tends to draw the fishnet knees wrong.
ハイレスフィックス使う時、網タイツの膝がぐちゃぐちゃにされるかもです。
v5.0: Underwear shape isn't consistent. Mix and match the panties prompts, and try with or without the prompt "thong".
下着の形がよく変わってます。「panties」のいろんなタグを混ぜて実験してみてください。「thong」の単語もつけてみてください。消すほうが良い場合もあります。
v5.0: Corset is sometimes drawn too low. In this case, use the prompt "underbust".
コルセットは時々低すぎるように描かれます。その場合、「underbust」の単語を使ってみてください。
Sometimes colors are drawn incorrectly. In v5.0, not specifying colors seems to draw them correctly by default. Try adding colors to the prompt if they do not.
時々色が違うように描かれます。v5.0では、色の単語を使わなくても正しく表示される場合が多いっぽいですが、そうじゃない場合には色の単語をつけてみてください。
Models used for screenshots （スクショに使ったモデル）:
I also used Adetailer and FreeU for some of the screenshots.
あるスクショにはAdetailerとFreeUも使いました。
AutismMix Pony
Hassaku v1.3
SweetMix Fp16Cleaned 2.1
MeinaPastel v5
CalicoMix v7.5
SeizaMix v2
YabaL Mix v5
ExpMixLine v3
Pushing Everything Off The Shelf Into The Pot v0.4
BeautyFool v1.5
SakushiMix v2.0
helloFlatCute2D v15e
MeinaHentai v4
MeinaPastel v6
Seen Anime
Description
FAQ
Comments (10)
Thanks for doing this, she's awesome ❤️ How did you train it with one picture: you cut the same image in different areas, I don't understand what you mean? Did you use the same crop size each time? I'm asking because I'm actually trying to train a character that I've created but I've drawn her only on one picture, so I'm interested in how you did this.
I took the original microskirt girl image, made the background white, then IMG2IMG upscaled it to get a big resolution version with cleaner details (describe some stuff so it knows how to fill in the detail). Then I made a 1024x1024 canvas and pasted/resized multiple copies of the new big resolution version, so as to frame important details. In other words, make zoomed in pics of parts I want the AI to recognize.
For example, one image is the full body image. Another is the face down to the collarbone. Another image is the waist up. Another image is just the calves down to the shoes. Another is the waist down to the shoes. A trick I learned from someone's guide on CivitAI is to rotate copies of the same image, as it seems to help the AI recognize details better. So I have two different images of the lower legs to the shoes, but rotated at different angles. I also suspect you can weaken the style by having copies of the same image but in different styles (use IMG2IMG to change style) but I haven't experimented with that yet.
For LoRA training I'm using AdamW8bit but unlike my previous LoRAs I tried something new. I used the info at the bottom of rentry org LazyTrainingGuide and also the new parameters Scale Weight norms at 1, Network dropout at 0.25, Rank dropout at 0.2 and Module dropout at 0.2. These seem to help with preventing overfitting. I also have the dataset folder named as "1_MiniskirtGirl", max train steps 2000 and 50 Epochs at train batch size 2. I didn't use Dadaptation3 because I haven't figured it out yet.
Oh, and I trained at 768x768. I used to train at 1024x1024, and I think it's nice to just keep the dataset large because it can always be resized down later.
Because it's just one image in the same one pose, when you use the LoRA it'll tend to draw the character in the same pose, but if you reduce the weight it becomes more malleable, at the cost of losing finer/specific details. Depending on what those details are, you can fill those back in via prompts. You can also just use full LoRA weight and then try overriding the LoRA by increasing the weight of prompts you use, eg. write (bare legs:1.5). Hope this info helps.
@Obliqeray Thank you for this complete explanation! Did you write specific tags for each of your images when training?
@Obliqeray very cool that you picked up a generated image to build a whole character out of it! @Gaspatcher thanks for pointing this lora to me :)
There are different approaches you can take to create a lora out of one image. You can artificially increase the dataset by doing what @Obliqeray suggested, or by getting more synthetic data via large batch prompting and/or controlnet, depending on what the subject matter is.
Another way to do it is to use a detailed txt with tags, and very low LR like even 0.00001, together with some other parameters suggested in this post, and I'd add multires noise at 0.1 and min snr gamma 5 as well.
One suggestion to you @Obliqeray would be to instead of asking the user to use lbw to limit output to MIDD in the lora prompt, you can just set the blocks outside of MIDD to 0 before training to achieve basically the same result. Assuming those blocks are the most effective to prevent affecting style when their weight is reduced.
@Gaspatcher after achieving somewhat good results you can just batch generate images again and feed back into another model.
When using synthetic data, besides tweaking block weights, a good idea is to use different base models and style loras to achieve a diverse style of synthetic outputs.
@Gaspatcher I didn't, I just used WD14 and then removed the tags that defined the character (eg. black hair), but left stuff like handbag (not part of base outfit) and standing, full body (part of composition). I just realized I failed to remove sleeveless shirt, but that's easy to override. I'm probably not tagging correctly.
As psoft said, you can always start with a one image LoRA and then use that to generate more images to increase the dataset, make another LoRA, and repeat. Any image details you're not happy with can be fixed with inpainting or photoshop editing, then back into IMG2IMG for clean-up.
One tip that comes to mind: if you're generating dataset images of a character with long hair covering their back, the AI tends to draw hair pulled over the shoulder instead of hanging down over the back. If you only use hair-pulled-over-shoulder images for the dataset, the LoRA won't be able to draw hair hanging over the back at all. It treats it like clothing. In that case, you'd need to disable the LoRA to inpaint the back hair, then once you're satisfied with that, use it as an image in your dataset.
@psoft Thanks for the info! I'd seen people talk about editing blocks but didn't realize it could be used that way.
@Obliqeray @psoft Awesome guys! Kudos to all of you for making me happy with you creations and teaching me how to be happy by creating my own! ❤️
@Gaspatcher If someone want to try, I have a quest!
Time ago, user aji1 got a wonderfull face by random. We discussed if there was a way to replicate it, but no luck.
We have just to images here:
https://civitai.com/images/1743044?period=AllTime&sort=Newest&view=categories&modelVersionId=26258&modelId=21990&postId=434683
Would be possible to make a LORA, or also a Textual Inversion just for the face?
@settima_ai Should be possible to make a LoRA based on one image of the face, but I dunno how good it will turn out. There's also this, which I saw a few days ago: https://github.com/modelscope/facechain but I haven't tried it myself yet.
@Obliqeray Not one... Two images! It's a rich situation, isn't it!?
@settima_ai https://youtu.be/SacK9tMVNUA