Melty || Shield Hero

Hi there, this page will contain all the information you need about my Melty lora.

This lora is intended to help in the generation of melty from "the rising of the shield hero" (aka "tate no yuusha no noriagari") and it was trained for 2 main reasons:

• the lack of understanding of the tag "melty_q_melromarc" by illustriousXL

• my intention of having a bunch of loras of the serie that were all at the same level.

This is a character lora, so the following sections will be structured as follows:

• A section containing the technical stuff.

• A section containing the tips and tricks.

• A section containing some tests.

• A brief summary.

Faq: Does this model support nsfw?

Short answer: Yes, but please don't.

This model was trained on a dataset of 25 images of the highest quality available. The dataset was not conditioned by any data augmentation tecnique, and the training went on for about 11 epochs with 4 repeats each.

The Lora is intended to have 1 activation keyword, which will have a general feeling of the character and that will stabilize the tags "blue hair, twintails, purple eyes" to make them reproduce the likeness of melty.

The activation keyword is:

m3lty

extra strongly suggested prompt:

blue hair, twintails, purple eyes

This lora is able to reproduce the Official outfit of the character, using the prompt:

blue dress, frills, white skirt

This lora has a little understanding of the anime style, so in some rare cases the tag "anime screenshot" might help reproduce the artstyle of the serie.

The model was trained on images that were mostly in the "landscape" resolution category, so it tends to perfomr better on orizontal images, but it has no major issues on vertical ones.

Due to the dataset, it is suggested to use the tag: "close-up" in the negative if it isn't needed.

The lora has no major bias conditions, but in some particular cases it might struggle a bit with some concepts.

It is the case of "v_sign" tag, that sadly reproduce most of the time "double_v_sign" instead of the single one; To fix this just place double_v_sign in the negative and tweak the steps just a little.

if you get annoyed about the hairbow being inconsistents you can stabilize them using the tag "hairbow" which is not required, but helps in obtaining more consistent images.

if you need help with the general prompt, keeping in mind that i am not really a good prompter, you might try something like:

masterpiece, high quality, best quality, anime screenshot, <lora:m3lty>
1girl, solo, m3lty, blue hair, twintails, purple eyes,
{attire tags}, {actions and/or poses}, {indoors || outdoors}, {background description}

It's test times:

The test were all done using hassaku illustriousXl with the following settings:

steps: 28~32
cfg: ~6

when used adetailer:
model: ultrasharp
steps: 15
denoising: 0.25

• expressions test:

  it seems to support the whole range of emotions.

• Clothing test:

  

  • the 1st image is the official outfit

  • the 2nd image is a random outfit (i've chosen the classic school uniform)

  • the 3rd test shoul've been the ???? version, but there won't be any for this model, for obvious reasons.

So what are my toughts on this?

I am happy with the results, it's pretty flexible, it didn't cost me too much (i am talking about the time and resources needed), and even toh the dataset was kinda bad, the result is pretty good.

This lora is part of my character loras collection you can find those in:

my profile -> collections.

faq: Will you train other characters from shield hero?

And at the time of writing i've already trained and tested few more loras from shield hero characters (malty, fitoria, mirellia), and those will be available pretty soon (maybe i'll publish them along with this one).
But considering that i am not intentioned to train loras for filo and raphtalia i guess i will stop (atleast for now) with those four characters. Maybe i will train glass and the other heroes in the future (but not any time soon).

That said, as always, i thank you for reading all this.

Remember if you like my work you can help me by following this account, publishing your content here, and share your works with your friends.

I hope you will find this helpfull and that i will see you for my next loras.