Pony version HERE.

This lora is trained on images produced from 2 different datasets, the first one combines images of 'Nicki Valentina Rose' and various bimbo images on the web. While the second is a merge of a 90s brunette pornstars with models from the black tape project.

Version 1 was trained using civit ai lora trainer

Version 2 was trained using google colab

the two versions trained on the same datasets, the intention was to make version 1 easier on weak rigs, while let version 2 be as normal as other loras for flux.

Trigger word: Apricota (using 'photo of apricota' helps).
additional triggers: bimbo, massive breasts.

Best between 0.7 ~ 1

Training Config: for version 1

unetLR = 0.0005
clipSkip = 1
networkDim = 2
numRepeats = 20
resolution = 1024
lrScheduler = cosine
minSnrGamma = 5
noiseOffset = 0.1
targetSteps = 4800
enableBucket = true
networkAlpha = 16
optimizerType = Prodigy
textEncoderLR = 0
maxTrainEpochs = 8
shuffleCaption = false
trainBatchSize = 1
lrSchedulerNumCycles = 3

Training Config: for version 2

precision to save = float16
Training steps = 4000
Epochs = 16
caption dropout rate = 0.05
optimizer = adamw8bit
Learning rate = 4e-4
Batch size = 1
use ema = True
ema decay = 0.99
linear = 16
linear alpha = 16

one of the example images is not using the natural language prompting, and it seems to work.

I tested this lora with flux1DevV1V2Flux1_flux1DevBNBNF4V2 and flux1DevV1V2Flux1_flux1SchnellBNBNF4

I used Forge Ui while testing and it worked fine with both.

--if you are having issues making loras works locally in forge ui (with flux nf4)

check this >link<

>BUZZ for cool images<

Have fun with Apricota