Karasu Chan - Artist Style 🎨
This LoRA is trained on the unique artstyle of Karasu Chan , known for their artworks about monster girls, you can checkout their creations on PixivFanbox or Pixiv or Twitter [contains explicit content].
-style mentioned by @ComradePedro-
Usage Tips :
regular use is at the weight of 0.8 to 1.1
No need for activation word but adding
karasuchanto the prompt will force the image elements toward subjects (NSFW) covered by the artist , and also gives the LoRA a bit more influence but might break other LoRAs.the LoRA is made to be flexible to work well with color ranges and detail levels, BUT if you're interested in the stylized approach of the artist u can add
monochromeorsketchor both 😋if you like that faint color glow that the artist has in most of his sketch you can use {color of your choice}+taint example:
monochrome, red taint
Fun Tips :
try exploring extreme expressions like
evil grincrazy laughextreme angerenvyloves to show teeth so try
fangs,shark teeth,clenched teethmix and mash monster parts like
horns,lizard tail,tentacle arms,spider arms. . etc, and see what you can create 😋
Samples settings :
Checkpoints: Anything V4.5, A-Zovya RPG Artist Tools V2
Steps : 24-26
Sampler : DPM++ 2M
ClipSkip : 2
have fun and consider rating ⭐
Description
FAQ
Comments (9)
looks interessting, i will give it a try later
also i've 2 questions cause i'll start making LoRAs/Lycoris too
1. do you train with images in original resolution or do you resize to like 512x512?
2. how do you prompt when training an art style like this?
thanks in advance ^^
no need to resize cuz the training algo will use multiple buckets to cover all resolutions.
I only edit the image to remove (Crop or Fill) elements that might create confusion when training, or scale down an unnecessarily large image to a size that doesn't alter needed details.
the second question is not really clear, do u mean prompt for testing ??
@navimixu for the second question i mean the text file that you need to create and put together with the image (in the training folder). where you have to write the prompts that describe the picture and tell the ai what it already had to know
and sorry for my bad english, i'm from germany and not used to speak/write english every day 😅
But thanks for the first tip and answers
@Enno89 that's called dataset captioning, there where the real work take place, with good captioning could even save a bad dataset, now where it comes to styles you need to describe the elements in the provided image in detail so that the ai focus on learning how they're illustrated instead of learning them.
lets say for example the artist im training a LoRA on is know for mostly drawing Cars, in order to just extract his style u need to mention that this is 'an image of a car in the style of ARTIST' that way it will try to learn how the car is illustrated in the image instead of fact that the image has a car.
usually you dont have to start the captioning from scratch you can use Interrogators like BLIP to skip like 80% of the work form there its just a bit clean of cleaning and altering. and you're ready.
A bit of trial and error and you'll get it right, good luck 👍
@navimixu Thanks, sounds like one big missing piece in my puzzle here 😊
now i'm curious how my next attempt will end and I'm also waiting for your next style LoRA, because I know and love a handful of the artists you've trained (like Geewhy, Tottotonero, Fluffydango, Mdf_an and Neocoill) 👍
sorry when i bother you with it but there are 3 last questions.
1. How many steps do you train per image?
I heard in a YouTube video that usally 100 steps should be enough. (for that you would name the image folder like "100_geewhy")
2. Is it the same to train on manga page images (that are only black and white)?
3. It seems that BLIP is disabled. Do you have a link for a working version beside the one on github (or do you run a captioning software localy on your system?)
@Enno89 no worries always happy to assist 👍
-It depends on the dataset length and learning-rate, the name ""100_geewhy" is specifying that each image in dataset "geewhy" 100 should have 100 steps so it that dataset have like 42 images then thats a total of (100x42) 4200 steps , overall in styles I aim to have around 2800 to 4000 total steps depending on the complexity of the subject.
-training on manga pages could be quite tricky, because it contains multiple frames that contains multiple scenes, and for that you need to take some time splitting each scene into separate image and cleaning the texts and line effects (its always better to have a clear understandable dataset then a large one), as for the monochrome colors dont worry if you specify it in captioning the ai will pick up on that and focus on the style rather then colors (meaning you could have colors in ur final results)
-most Colab Trainers includes an implementation of BLIP to work on the dataset directly from drive, I sometime do them locally in auto1111 if I feel the need to for alot of manual changes then upload them back to drive.
Yessssss! Thank you! It looks so good! You are a real homie man, cheers!
u have my thanks as well , didn't know about this amazing artist till now xD
@navimixu Hello there again! I was thinking, if you wouldn't mind. There's a specific thing that Karasu_chan does that i'd like to replicate as well, which is basically pitch black hoodie with glowing eyes and tentacles under the skirt. I'd really appreciate you doing that, thank youuu! (Or you can make those a separate lora, your call.)
Here's also a site with all of their work they've done so far: (BEWARE, NSFW PRESENT!) https://e-hentai.org/tag/artist:karasu+chan
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.