First step of my LoRA training for this OC: making embeddings to help generate her "en masse".

• EleanorDescILL (56 tokens): Her hair, eyes and makeup, together with her mechanical arms

• EleanorOutfitILL (37 tokens): Her standard outfit including footwear

Sample prompt:

1girl, solo, smiling, EleanorDescILL, interesting angle, eye-catching, crossed legs
BREAK
EleanorOutfitILL, simple background, black background, MechaP0s

NB: i am posting those here only to describe and link pictures using them. They probably will not be used for long 😅

Long Story

When doing those embeddings, i am using Embedding merge. First step is sanitazing the original prompt:

1girl, solo, cowboy shot, dynamic pose, smug, (orange hair):1.2, short hair, hime cut, (blue eyes, black sclera), black lipstick, makeup, eyeliner, mascara, (bright green shirt, buttoned up, gold buttons), sleeveless, suspenders, (full mechanical arms, mechanical shoulders, gold metal, gold robot joints):1.1, black shorts, leather shorts, garter straps, (combat boots, black boots, punk boots), simple background, black background, MechaP0s

Becomes:

(orange hair, short hair, hime cut), orange hair, (blue eyes, black sclera), black lipstick, (makeup, eyeliner, mascara), (full mechanical arms, mechanical arms, mechanical shoulders, gold metal, gold robot joints)

and:

(bright green shirt, buttoned up, gold buttons), sleeveless, suspenders, (black shorts, leather shorts), garter straps, (combat boots, black boots, punk boots)

The goal is to remove any attention (the :1.2 and :1.1), group stuff by "Description" and "Outfit", duplicate a few tags for what may be skipped (hair color and the mechanical arms here) and "join" using parenthesis elements that belong together.

If increase the weight is needed, you supposedly can multiple tags using EM, but I found it does not works perfectly. I usually do this afterwards manually (for example):

>>> from safetensors.torch import load_file, save_file
>>> o = load_file("EleanorDescILLbefore.safetensors")
>>> n = {}
>>> n['clip_g'] = o['clip_g'] * 1.2
>>> n['clip_l'] = o['clip_l'] * 1.2
>>> save_file(n, "EleanorDescILLafter.safetensors")
>>> exit()

The next steps are doing a few tests, adjust if needed (be aware restarting Forge/A1111 is needed when deleting/remaking embeddings with the same name).

Now, onward to masse generation 😁

PS: those embeddings don't work on-site and were made using NewMecha. They may not work as well with other checkpoints.

Addendum: i did a small patch on the description embedding using a small trick: simply renaming and reusing the previous version with additional tags:

EleanorDescILLold, blue eyes, robot arms

The eyes sometimes ended up orange and depending on the pose both arms weren't mechanical. This is still not enough for consistancy, but good enough before ensuring this via a LoRA.