Illustrious(ILXL): Trained on Illustrious-XL(v0.1).
NoobAI-XL(NAI-XL): Trained on NoobAI-XL EPS1.1.
null-ghost:
https://www.furaffinity.net/user/null-ghost
https://twitter.com/NullGhostArt
https://www.patreon.com/NullGhostArt
Optional Prompt :
2024, 2023, 2022, 2021, 2020, 2017…Prompts for ILXL
Prompt Prefix:
masterpiece, best quality, amazing quality, Negative Prompt:
bad quality, worst quality, worst detail, sketch, censor, Optional Negative Prompt:
monochrome, greyscale, signature, url, artist name, username, watermark, text, English text, Prompts for NAI-XL
Prompt Prefix:
masterpiece, best quality, newest, absurdres, highres,Negative Prompt:
worst quality, old, early, low quality, lowres, signature, username, logo, bad hands, mutated hands, Description
{
"engine": "kohya",
"unetLR": 1,
"clipSkip": 2,
"loraType": "lora",
"keepTokens": 1,
"networkDim": 32,
"numRepeats": 2,
"resolution": 1024,
"lrScheduler": "cosine",
"minSnrGamma": 5,
"noiseOffset": 0.1,
"targetSteps": 3510,
"enableBucket": true,
"networkAlpha": 16,
"optimizerType": "Prodigy",
"textEncoderLR": 1,
"maxTrainEpochs": 30,
"shuffleCaption": true,
"trainBatchSize": 6,
"flipAugmentation": false,
"lrSchedulerNumCycles": 3
}
FAQ
Comments (6)
Amazing, I am deeply grateful for your selfless and outstanding work as well as my favorite artist null-ghost and those masterpieces made by null-ghost.
Though this lora is in testing but it seems not having Big issues, May no need further enhancement but I will greatly appreciate it if you willing to!
My simple testing:
Lora weight=0.9 is ok, stylish enough without breaking image.
Suggest using species tags you want on e621 like "canid,felid,crocodile...... Etc." in prompt as assistance (Which null-ghost drawn most and those images get tagged).
Using Only "null-ghost" without "2024,2023,2022...... Etc." In prompts is totally ok (and the images are even more creative?)
2024,2023=more smooth and cinematic?
2021,2020,2017=more colorful?
2015,2014,2013,2012,2011,2010=retro Japanese anime? (just as those old artworks by null-ghost)
Little issues:
1. When generate some less popular species (like "wild boar"), It seems that it will decrease "null-ghost style" very much and lowering image quality slightly. (fair enough, null-ghost seldom draws them)
2. When generate distant view, the characters' face get a little bit distortion, not too bad but not as perfect as "sharp focus" ones (especially the eyes). Or maybe it's because of the checkpoint I used Or my bad prompts?
Anyways, expressing my gratitude to the selfless hardworking author and the extraordinary artist null-ghost once again.
Thank you for your usage and feedback! Tags like "2024" are used to subtly differentiate styles, as an artist's style can evolve over time. Although many images lack year tags, the illustrious model's nature means that altering a single tag can significantly impact the image.
Due to the canvas size limitations of the website generator, the full potential of the checkpoint cannot be fully realized. Additionally, there are many excellent plugins, such as those that enhance eye and body details. If the model could be trained locally or run on a local machine, the results would likely be even better. Thank you once again for your comments!❤️
May I ask which base model are you using to train IL LORA,What is 889818.safetensors？
For compatibility, I use illustrious xl.
https://civitai.com/models/795765/illustrious-xl
By the way, where did you see the number 889818? The first item in Google search for this file name is also the illustrious xl I used.
@sxus_Sw Thx,I saw that number in webui Lora infomation,I searched it in civitai and got nothing.Maybe next time I should try Google first.
@morenatsu502 Thank you, find answer thankfuly to your question and 889818, lol.)