(now for me,, free user ,, the best result from this lora cgf 20 sample 20 )

v.1 is the 8 epo .. v.2 is the last epo i train ( epo 14 )

so v.3 will be the epo 13 ,v4 will be the epo 12 and so one

i use caption in this training , 1qq trained without any caption

i use the same dataset as 1qq lora ( from the example image ,, i have 3 women face in this lora )

i look and i look again ,all of them are likely original flux women ,, i need more work here..

no problem, clicking and testing is fun anyway

there will be 3qq soon lol

for this 2 qq ,,i keep 8 ,9,10 .....----> final epo

2qq have many version for testing now lol (8 version planing )

( i use trigger too ..2qq )

setting .. i use 35 image with 50-150 shapeness , i create images from ai

"engine": "kohya", "unetLR": 0.00005, "clipSkip": 1, "loraType": "lora", "keepTokens": 0, "networkDim": 64, "numRepeats": 12, "resolution": 1024, "lrScheduler": "cosine_with_restarts", "minSnrGamma": 5, "noiseOffset": 0.1, "targetSteps": 2940, "enableBucket": true, "networkAlpha": 32, "optimizerType": "AdamW8Bit", "textEncoderLR": 0, "maxTrainEpochs": 14, "shuffleCaption": false, "trainBatchSize": 2, "flipAugmentation": false, "lrSchedulerNumCycles": 3

200 step /epo , may be its too low ,, lets have fun,, i use this website for training and i use this website for testing it too,, if you want to test it then i say thank you