Several (like a lot) of pictures of this very sexy Goth Girl were posted by @coolhotrod813 on New Mecha Gallery, enough to make me want to have a LoRA of her! 🥰

And since i wanted to experiment a new setting, here we are! Happy New Year 🥳

THIS OC IS BY @coolhotrod813, if you like it, don't tip the model, tip the original creator 💜

Trigger (a bit long and descriptive for a change):

C00lH0tG0th, 1girl, black hair, choker, large breasts, gothic girl, dark eyes, dark eyeshadow, black lipstick, mascara, eyeliner, eyelashes, high twintails, skull hair ornament

If you check the showcase picture, you'll notice punkG0th which is a temporary embedding i used to describe her basic outfit:

black shirt, cleavage, black corset, pencil skirt, short skirt, black skirt, fishnet stockings, high heels

The training settings

I wanted to test a smaller dimension for a change, with a 1/4 ration alpha to dim.

{
  "engine": "kohya",
  "unetLR": 0.0002,
  "clipSkip": 2,
  "loraType": "lora",
  "keepTokens": 1,
  "networkDim": 16,
  "numRepeats": 2,
  "resolution": 1024,
  "lrScheduler": "cosine",
  "minSnrGamma": 0,
  "noiseOffset": 0.03,
  "targetSteps": 1420,
  "enableBucket": true,
  "networkAlpha": 4,
  "optimizerType": "Prodigy",
  "textEncoderLR": 0.00002,
  "maxTrainEpochs": 10,
  "shuffleCaption": true,
  "trainBatchSize": 1,
  "flipAugmentation": false,
  "lrSchedulerNumCycles": 1
}

Tagging was done using WD14 moat tagger v2 and some very small adjustment (essentially, removing some tags).

Training was done against Neural Lens Core.

The sd-mecha test

I wanted to test the N-average merge feature of sd-mecha, feeling it could be a great way to get something easily out of several epochs. From my training (done on-site), i took epoch 1, 2, 4, 5, 9 and 10 and then:

from glob import glob
import sd_mecha
sd_mecha.set_log_level()

models = glob("*.safetensors")
models = list(map(sd_mecha.model,models))

recipe = sd_mecha.n_average(*models)
sd_mecha.merge(recipe, output="C00lH0tG0th.safetensors")

Simple as that, i took all epoch from the folder where they were, mixed them all together and got this LoRA 😁

PS: 3/4 of the showcase pictures are flagged for review 😣

PS2: OK, got it, it was because i am using an embedding that is not published /o/ Well, let's publish it then 😋

PS3: I have added the resource manually but i will probably have to delete/re-upload the corresponding pictures 😭