update: for Animagine XL 3.1
notice: if animagine 3.1 have tooooooo black outline,use my special outline lora with negative weight. https://civarchive.com/models/169170?modelVersionId=383532
recommend weight:0.8~1
none Trigger word.
beacuse base model include some ba style,if use "blue archive" or other character name from ba in prompts, need a low weight of this lora.
other use it normal
This lora train use Blue Archive Game CG base Kohaku-XL Delta ,and I have not test it with other base models, so it should not be usable.
after also training a ver base on pony/other model. However, I cannot sure other ver is enough stronger as it has only basic testing.
Use the same License like Kohaku-XL Delta.
License
This model is released under Fair-AI-Public-License-1.0-SD
Plz check this website for more information:
Freedom of Development (freedevproject.org)
Description
FAQ
Comments (4)
大佬，可否做一个适用pony版本的，您的aurora是我用过最好的pony模型，不要放弃pony😭
....aurora应该不会再更新了,那个我炼得很痛苦,最后是用4万张图跑到overtrain再mbw往回调的..无论对算力还是精神的损耗都太高了....aurora算上测试前后花了接近20天4090满负荷运行，这个时间损耗已经足够我用1m图给kohaku继续训练几个ep了，而且人会轻松很多。
包括这个在内的以及后续的画风lora我应该都会补一个pony版本，但是测试应该不会像之前那样覆盖那么多了。
@kitarz 能请教下pony训练的工作流吗？我试着用BAXL的数据集在上边微调，效果很差……不太清楚是tag的问题还是什么
@4BEraser 没有什么具体的工作流，大概就一些经验？ 1.不要动pony的te。 2.如果吃了某些lora/融合了任何模型动了te，最后一步一定要把pony的te的给替换回来。 3. lr开大训练久一点 4.caption可以用一下a3的caption规则，但是我不确定是对的，我只是最后测试到了这边。5.如果不打算用足够量的训练集训练足够长的时间的话，只是覆盖一个画风我建议是在pony上训练，然后合并差异/训练的lora/lokr 到aurora或者”js2pony替换te“上合并效果应该会好很多。
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.