[anima base1]

anima的lora的画风还原程度貌似很难达到像sdxl那样可以直出原图，并且部分细节几乎无法还原，使用了更激进的学习率也只能做到8成像，尽管其他方面都要优于sdxl，但还是挺难以接受的，所以还在找其他方法。

The LoRA for anima seems to have a hard time matching the style as accurately as SDXL, which can reproduce the original image directly. Some details are almost impossible to fully recreate. Even with a more aggressive learning rate, it only gets about 80% of the way there. Although it's better than SDXL in other ways, this is still hard to accept, so I'm looking for other methods.

强度Strength：1-1.2 触发词Trigger words：@1nnerkey, 美露莘风格Melusine style：melusine \(genshin impact\), furry female, furry, body fur, animal ears, animal nose

如果图片出现水印和日期，在负面提示词加上：signature, dated

If watermarks or dates appear in the image, add this to the negative prompt: signature, dated

此外，我修改了我的工作流，导致正面提示词和使用的lora无法被正常读取，所以你需要将图片下载下来以读取它们。

Also, I've modified my workflow, so the positive prompts and the LoRAs I'm using can't be read properly. That means you'll need to download the image to read them.

[anima preview3]

最近尝试了在anima preview3上进行lora训练，与noob vpred1.0相比，anima的lora要优于noob，但训练时长是SDXL的两倍多。lora有些欠拟合，可能需要提高Lora强度。

Lora强度为1~1.3。触发词：@1nnerkey

如果图片出现水印和日期，在负面提示词加上：signature, dated,

anima发布正式版后，会重新训练。

Recently I tried training a LoRA on Anima Preview 3. Compared to Noob VPred 1.0, Anima's LoRA gives better results, but training takes more than twice as long as SDXL. The LoRA is a bit underfitted, so you may need to increase the LoRA strength.

strength: 1-1.3 . trigger word：@1nnerkey

If you see watermarks or dates in the image, add "signature, dated" to the negative prompt.

Once the official version of Anima is released, I will retrain the LoRA.

[noobai vpred1.0]

一次实验性测试，但训练出来的LoRA效果我觉得是完全可用的状态，所以我把它放了出来。
触发词：1nnerkey

An experimental test, but the LoRA trained from it is, in my opinion, fully functional, so I've released it.
Trigger word: 1nnerkey