Experimenting with training a character LoRA starting from one image. Inspired by https://civarchive.com/articles/3021/one-image-is-all-you-need
{
"unetLR": 0.0005,
"clipSkip": 1,
"loraType": "lora",
"keepTokens": 1,
"networkDim": 64,
"numRepeats": 20,
"resolution": 1024,
"lrScheduler": "cosine",
"minSnrGamma": 5,
"noiseOffset": 0.1,
"targetSteps": 1400,
"enableBucket": true,
"networkAlpha": 32,
"optimizerType": "Prodigy",
"textEncoderLR": 0.00005,
"maxTrainEpochs": 20,
"shuffleCaption": true,
"trainBatchSize": 8,
"flipAugmentation": false,
"lrSchedulerNumCycles": 3
}
Description
Trained on images from v1.0. Theoretically more robust and flexible.
FAQ
Comments (5)
Is the one picture in preview the one that the lora was trainer from? (Or can you post or let us see the one that was used?)
Here you go: https://civitai.com/posts/6850425
For v1 I took that initial image and cropped and scaled it to make a 28 image dataset with a variety of image dimensions and aspect ratios
@WhiskyBoat Interesting. Thank you. It is only 1/4 body-shot. That means the body is generated more from model standards. At least, that's what I would think. What was the base model for training this lora?
@HugMeIntoFace For v1.0 that is certainly the case; the body is pretty much entirely determined by the checkpoint you use, other than a propensity for white t-shirts with a print. v2.0 had some 3/4-shots in the dataset so that version has more influence below the shoulders.
v1.0 was trained on-site with 2dn-pony (v1.0), which was the model used to make the original image. v2.0 was trained on-site with Pony, because I was too cheap to spend twice as much buzz.
@WhiskyBoat Thanks for explanation!! It is interesting. The v2.0 one was created from the wider shot, but it seems that the face features was kept intact. (Maybe I will try to train from one picture in the future or from two (front/back) if that would be any wise...)
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.