updated, added two more resources in the list

Introduction

I trained two LoRAs using Kohya_ss and really, i didn't know what i was doing.

I vaguely followed this tutorial and achieved "something", but this is less that perfect. Since i want to integrate "BancinXL" style in an Illustrious LoRA, i need to know better.

So, for a first try in re-learning properly how to use Kohya_ss, i'll do it in two steps:

• Generate a new LoRA for Pony and validate the process,

• Generate with the same dataset a LoRA for Illustrious and check the result.

• Optional: try it with a v-pred model.

To do so, i'll leverage as much other tutorials and resources i can:

• https://civitai.com/articles/3522/valstrixs-crash-course-guide-to-lora-and-lycoris-training

• https://civitai.com/articles/4100/how-to-create-a-style-lycoris-for-ponyxl-on-low-vram-locally

• https://civitai.com/models/1006542/lora-training-local-steps-calculator

• https://github.com/bmaltais/kohya_ss/wiki/LoRA-training-parameters

• https://civitai.com/articles/9585/joscheks-lora-training-tipps

• https://civitai.com/articles/138/making-a-lora-is-like-baking-a-cake

• https://civitai.com/articles/6173/demystifying-snr-min-snr-debiased-estimation-and-ip-noise-gamma

• https://civitai.com/articles/6264/a-simple-script-for-visualizing-captions-in-large-datasets

• https://civitai.com/articles/7766/resourcestools-and-tricks-i-use-for-outfit-loras

• https://civitai.com/articles/6990/guide-on-lora-making-with-multiple-outfits

• https://civitai.com/articles/3342/regularization-from-the-ai-layman-perspective

• https://civitai.com/articles/9148/illustrious-lora-training-discussion

• https://civitai.com/articles/11394

• https://civitai.com/articles/83/using-tensorboard-to-analyze-training-data-and-create-better-models

• https://civitai.com/articles/5545/pony-realism-lora-training-and-preset

• https://cameronrwolfe.substack.com/p/the-best-learning-rate-schedules

• https://civitai.com/articles/680/fictional-characters-among-other-things-a-lora-training-guide

• ... (i'll add here whatever other things i may find)

But the first step is... what should i train my LoRA on!

The idea

I used to play WoW a lot in the old days, and my Warlock Liadora could be a good target for an OC. And it could also be a good idea to test other LoRA/Checkpoint in the future (and it create an identity to my content).

The character had short red hair, elf ears (once i re-rolled from humans to blood-elves) and i was sporting my Felheart Horns everywhere, even after getting better gear. And black stuff, always black stuff :D

To go for that, i'll have to generate a set of picture with the intended look and feel. That's why i created the g0thOutfit embedding, together with an other one (unreleased, it would not be useful for other people): LiadoraDesc

This positive embedding, like g0thOutfit, is just a token merge for CLIP, based on this:

((Yorha 2B, red hair, green eyes, dark red lipstick), (elf ears, ear piercing), (oni horns, black horns))

This allows me to create more simple prompts for generation and keep a bit of consistency.

I cheated by using 2B as one of the token to get the haircut and features i wanted and altered the other features (hair and eyes color, extra stuff...)

Together with g0thOutfit, this is a good starting point with a simple short prompt:

Expressiveh, PonyScores7, masterpiece, best quality, 1girl, solo, portrait, holding own hair, facing viewer, LiadoraDesc, smug, large breasts, latex g0thOutfit, cleavage, (purple gloves):1.5, simple background, volumetric lighting <lora:MM_Gothic_NJ_Mix:0.8>

Now, i'll generate around 150 pictures in various positions and try my hand on the parameters ;-)

Soon more info on if i managed to do it correctly and what i did in the end ;-)

(And i'll also change my avatar to a "Liadora" picture to keep me motivated :D)

Thanks for reading! <3