Modified Anima Yume for LoRA training.
It reduces unintended style changes.
This is not designed for the original model(anima-preview.safetensors). Plaese use Any Anima Preview instead.
Note: This model is NOT suitable for image generation. The model tends to generate ugly images.
LoRA学習用のAnima Yumeです。
意図しない画風の変化を軽減し、精度も改善します。
Dataset information:
Dataset size: 5120
Gelbooruから収集した5120枚の画像で学習しました。
なお、以下のタグがあるか条件を満たす画像は除外しました。
filetype:gif, score:<0, mpixels:<1048576, tagcount:<16, \*_artifacts, adversarial_noise, greyscale, monochrome, digimon, photophop_(meidum), ai-generated, duplicate, bad_\*, off-topic, cropped, resized, reversed, rotated, third-party_edit, screenshot, tagme, real_life, watermark, 3d, koikatsu_(medium), mikumikudance, twitter_username
Description
Target version: v0.4
FAQ
Comments (2)
Using xyz comparisons, I believe no one will clearly see the differences, so I'll just state my conclusion subjectively. Training lora: Preview2 > AnimaYume0.4; Generating image: AnimaYume0.4 > Preview2. Regarding AnimaYume0.4, its "white border" does disappear or narrow, and the aesthetics are indeed improved. However, with the same "training parameters," it didn't meet my expectations; I won't elaborate on why. The original author's goal was to generate images that better fit the "illustrious" aesthetic... I think if you only need to training character, you can use "AnimaYume0.4" as the training model; for other more challenging themes, please consider "Preview2". The LoRa models trained on these models are interchangeable and not mutually exclusive, as there is little difference between them. However, I believe Preview2 is more reliable in terms of training, while AnimaYume0.4 is more flexible in terms of image generation.
Oh thank you