WitchBlade XL

This is a Lora based on the character Sara Pezzini from the WitchBlade comics. While it was trained on many images from the comics, the dataset also includes higher detail cover art and realistic images to give it more flexibility.

The are now three versions, each with their own strengths and weaknesses. I don't have any plans to make other versions, but anyone is free to use any of these models for additional training/merging.

Version 1.0 was trained through CivitAI, so it's pretty simple. The captions and image quality weren't as good with this one, but it sticks closer to the original comic book style. I've tested it with a strength up to 1, sample steps down to 20, and a config scale up to 7 without any notable issues. However, it is not as versatile as v2, due to the limited dataset. (It can create male characters, but they will look like a masculine version of the Sara Pezzini character)

Version 2.0 was trained on my PC with a higher quality dataset and better captioning. The dataset for this version also includes male characters, so it does better with male prompts than v1. This dataset also included more realistic and painted styles, using high quality cover art and other images, so it has less of that comic book style, and works better for realism and detailed images.

However, it was overtrained. I didn't notice this in my testing because I didn't try to mix other styles as much, but after using it with Civit AI's image generator I noticed it heavily influenced the outputs. It works better at lower strengths of 0.2 - 0.5 with most other loras (but can go higher with certain loras), and the config scale seems to be best between 3-5, it may also require a larger step count of 50-60.

The final version is a merge between v1 and v2, with a slight lean toward v2. While it does have more of a comic book style than v2, while keeping some of the versatility, it still isn't as close to the original style as v1.

I do have some examples for prompts that may be of use. However, the following images and information below uses the v2 model, and you'll likely get different results with v1 and the merge. I updated the images for each version, using the same prompts and seeds for comparisons.

Note: All of the following examples were made with very simple prompts and low settings, so they are not the best quality, they are only used for comparisons. Also each example will include a artistic and realistic style.

First off, because a much larger number of images for the WitchBlade are of the main character, Sara Pezzini in the "bikini armor" style suit, both models lean heavily toward that design.

If you want to get the character without the armor, you should be able to do this my simply adding clothing to your prompts, and not including any "armor" in the prompt or even putting it in the negative prompt. I tried to also get this to work with the WitchBlade gauntlet, but it seems to have a higher chance of being added even if you don't specifically prompt for it.

Sara Pezzini often wears blue jeans, a shirt or tank top, and a leather jacket, so using this prompt we get the following; "WitchBlade, Sara Pezzini, gauntlet, jeans, t-shirt, leather jacket, comic book art/realistic"

She is also featured early on in an iconic red dress, which can be prompted by simply changing to; "WitchBlade, Sara Pezzini, gauntlet, red dress, comic book art/realistic"

I also tried to include the main two armor types, the "bikini armor" as mentioned above, and a full body armor, prompted like; "WitchBlade, Sara Pezzini, gauntlet, full armor, comic book art/realistic"

and; "WitchBlade, Sara Pezzini, gauntlet, bikini armor, comic book art/realistic"

As for the male characters, I didn't have enough high quality images for the dataset so I didn't bother using character names, instead it's just a simple "male WitchBlade" tag that replaces the "WitchBlade" and "Sara Pezzini" tags.

example 1; "male WitchBlade, gauntlet, pants, black shirt, comic book art/realistic"

and example 2; "male WitchBlade, gauntlet, full armor, comic book art/realistic"

All of my models may be used for other training and merges.

描述:

Better quality images and better captions in dataset.

Works best at mid-lower strength, cfg scale 3-5, and step counts above 40.

More painted detail or realism than comic style.

训练词语: WitchBlade