Recast:

Strictly speaking, this isn't a brand‑new version. Since I couldn't locate the complete training set from the last one, I ended up swapping out some of the source material, re‑doing the annotations, and making sure all the images in the training set share the same resolution. It might fix a few issues from the previous version (no promises, though—I'm not 100% sure), but it probably introduces a couple of new quirks as well. Unfortunately, the relevant assets are still far too scarce, so I honestly can't say whether a better version will ever see the light of day. As far as the output goes, this version tends to produce more extreme body proportions, which—for better or worse—does align with what I had in mind from the start. I'd still suggest treating it as an "accessory" rather than the main course.

Anyway, hope you have fun with it.

This is an experimental Lora. While using various Checkpoint models, I noticed that some characters' outfits and body types are often tied together. Sometimes when generating images of certain characters, I'd struggle with inaccurate outfits—because adding too many body-related prompts would interfere with the clothing. That's why I tried creating this extreme Lora. With some outfit-related prompts for the character, this Lora can give the figure an exaggerated look while trying to keep the outfit as intact as possible… well, in theory.

In practice, it’s not that simple. This Lora does affect the overall art style to some extent, and if you use "large"-related prompts for every body part without lowering the weight, it will generate extremely exaggerated proportions—which can be a real issue for online generation. So I’ve uploaded a relatively early version here, but if you decide to use it, I still recommend lowering the weight.

The reason behind these quirks is that I merged several different artist styles, mixed in some AI-generated images, and also added some of my own AI works. In a way, it’s also a reflection of my journey over the past two years.

Overall, I’d recommend using it more as a supporting concept Lora—it can also play a role when blending art styles. The training dataset was quite small, so this might be its only version.

Anyway, as always, I hope you have fun

I used a translation software to translate my words. Allow me to apologize if it has distorted my meaning.