Liadora (OC)

V1.1 update: The V1.0 was not correctly detected when an image was uploaded to civitai, probably due to a mismatch between the hash in the metadata and the real hash of the LoRA. Since i wanted to fix this, i took this opportunity to tweek the TE layers of the LoRA to avoid overriding the model style too much when using this LoRA (and this allowed me to redo some more showcases pictures using New Mecha Duchesse 💜)

PS: well, lora detection is still broken (or maybe it's on my side T_T)

My OC Liadora character LoRA!

🎉At long last, i managed to train a V1 that i like enough to publish. 💜

It is the result of several stars aligning:

• New Mecha release and its rendition of Liadora embeddings

• Neural Lens Core as a training base instead of whatever i tried before

• Better understanding of some training parameters (especially Min SNR)

• Cleaner dataset than previously, with special thanks to Diamond_Letter and C0rn_Fl4k3s wildcards (i will cleanup the resulting mixed wildcard file for release)

I have baked in her outfit as a specific tag (THIS is a valid use case for training the TE).

Triggers:

• Liadora,1girl,red hair,green eyes,black horns: the base description needed

• Liadora,1girl,red hair,green eyes,black horns,liadoraoutfit: baked in full outfit including the shoulder tattoos

Here is the training parameters used (with 120 pictures):

{
  "engine": "kohya",
  "unetLR": 0.0001,
  "clipSkip": 2,
  "loraType": "lora",
  "keepTokens": 0,
  "networkDim": 32,
  "numRepeats": 2,
  "resolution": 1024,
  "lrScheduler": "cosine",
  "minSnrGamma": 0,
  "noiseOffset": 0.03,
  "targetSteps": 1440,
  "enableBucket": true,
  "networkAlpha": 8,
  "optimizerType": "Prodigy",
  "textEncoderLR": 0.00001,
  "maxTrainEpochs": 12,
  "shuffleCaption": false,
  "trainBatchSize": 2,
  "flipAugmentation": false,
  "lrSchedulerNumCycles": 1
}

PS: Her look and style is heavily baked in the LoRA and will bleed on other checkpoint default style: that's an intended result as i really wanted her to look THIS way. 😅

A special thanks to everyone who participated in my bounties and motivated me to finally release a Liadora LoRA (you know who you are 💜)

Addendum: As an unexpected feature, the Liadora Outfit is usable as a standalone in theory 🤩