The Monocled Werewolf in game: Atelier Yumia: The Alchemist of Memories & the Envisioned Land
https://atelier.games/yumia/us/characters/villain01.html
I only used less than 10 pictures to train this time, so don't expect good results. Actually, I only have five usable pictures: https://civarchive.com/articles/9415/training-data-for-the-monocled-werewolf
Steps: For the first time, I suggest that it should be as few as possible, and you can even use draft mode.
Strength: 0.6～0.8, When hi-res fix, it can be reduced as appropriate.
A little problem is that I can't tell whether he has ears … I prefer that he doesn't have ears, but it's too difficult to control the ears from generating through negative prompt, so it doesn't matter. Let's give him a pair of lovely ears.
Please do not use the generated pictures in the official fan competition.
When I was training it, I turned on Flip Augmentation. Therefore, when generating pictures, the character's shoulder armor and glasses are sometimes mirrored, which is to expand the training pictures.Maybe I will update this question in the next version.
Description
{
"engine": "kohya",
"unetLR": 1,
"clipSkip": 2,
"loraType": "lora",
"keepTokens": 1,
"networkDim": 32,
"numRepeats": 12,
"resolution": 1024,
"lrScheduler": "cosine",
"minSnrGamma": 5,
"noiseOffset": 0.1,
"targetSteps": 1054,
"enableBucket": true,
"networkAlpha": 16,
"optimizerType": "Prodigy",
"textEncoderLR": 1,
"maxTrainEpochs": 31,
"shuffleCaption": true,
"trainBatchSize": 6,
"flipAugmentation": true,
"lrSchedulerNumCycles": 3
}
Comments (4)
I don't hate you in any way, on the contrary I adore your models, but why put up a low-quality model? You could have tested it, figured out what was wrong and put up a more developed and high-quality version.
Hello, thank you for your comment. In fact, I haven't conducted a complete test yet. At present, the model trained by the only five pictures I use basically meets my requirements, because there are so few training pictures! I can't expect better results.
Training data:https://civitai.com/articles/9415/training-data-for-the-monocled-werewolf
I don't think only these pictures can train a high-quality model. But I don't think it is of low quality. The unique Japanese style contained in the character is not as real as possible, and the Illustrious model is not for realism. To say the least, even if it is of low quality, it still has the value of being published. At low strength, it can be used as a decorative model. Finally, thank you again for your concern.If you have more pictures about this character, I think we can improve this model together.
When generating the image of this character, my practice is to use the draft mode first, and then use the hi-res fix. When using the hi-res fix, the strength of this model is reduced by about half, which can generate better details while retaining the characteristics of the character.
Thank you. Your suggestion is very helpful.
Details
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.