Test 18
Trigger words: y0k00zuma
I posted this model a few days ago, but deleted it due to an operational error.
Version F
Only suitable for generating images of a single girl. Its training dataset only includes paintings of single women by the artist. But it is the best at reproducing the artist's style.
Version 1.0 (the version posted last time)
Suitable for generating images containing men. The reproduction of the artist's style is slightly worse than version F (a very subtle difference, because its training dataset added paintings containing men and multiple characters to version F, so the portrayal of women is not as focused and extreme).
Version H
Represents the artist's black and white (or monochrome) print illustration style. The training dataset images come from illustrations in a novel called The Gates of Hell, which do not feature tattoos.
In fact, I also trained an Anima version of Lora, but the results were far from what I expected. I used several popular checkpoint models to generate images, but the results were either too cartoonish or resembled digital illustrations, failing to achieve the soft, delicate lines and watercolor texture.
Of course, this could very well be because I didn't set the parameters correctly. But in any case, I will not publish the Anima version of lora until I get a result that satisfies me.
Description
FAQ
Comments (10)
I'm interested in your art style.You should give ANIMA a try. If you're familiar with this kind of thing, I think you'll notice it has much better style rendering capabilities.
I’m not sure why the Anima LoRA I trained didn't turn out well, it might be because only half the images in the training dataset used natural language tags. I’ve only recently started working with Anima and am not yet very familiar with it, though it seems to require more detailed prompts.
@Aaaher777 Give ANIMA more time and experiment with it. Once you really get to know how it works, you'll find that many of the limitations you've run into with IL can be improved. It won't solve everything, but the improvement is real. Trust me.
艺术性很高
I saw your comparison between the Illustration and ANIMA results. Are you unhappy because the ANIMA version seems to lose some of the finer details and brushstroke-like artistic style? I wouldn't expect that, though. From what I've seen, ANIMA should generally be better at reproducing texture, especially when it comes to images with a strong artistic style. Or am I getting the wrong impression?
Yes, I think Anima image is too "clean."
I’ve finished training an Anima LoRA based on my sketches, but I’ll be uploading it a few days late, because it’s really hard to choose example images that I'm satisfied with.
Perhaps because I was training locally, I couldn't set the learning rate extremely high. However, the results were already much better than those of the Civitai trainer with its default settings, and it took twice as long as Illustration Lora to train.
@Aaaher777 Would increasing the network (Rank / Alpha values) help?
@BIG_A I'll try that next time. one training on the PC took me about 6~8 hours, it's difficult for me to train every day.
I'll test it with a few pictures first.
@Aaaher777 That's actually pretty strange. In theory, ANIMA should outperform Illustration when it comes to preserving details and reproducing artistic styles. But in reality, one of my friends has also complained that ANIMA drives him crazy when trying to reproduce traditional painting styles.
I'm planning to do some experimental training with a few style reference datasets on my own to figure out what's really going on.
Another headache is the training time. I usually start training right before I go to bed, and by the time I wake up, it's finished. Of course, one reason I can do that is because electricity is quite cheap where I live.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.