I used 22 photos for training. 2600 steps. I trained on the model - cyberrealistic-1.2. and the dpm++ 2M karras sampler. when generating the image itself, I use what is described above and Denoising strength 0.4 and CFG scale 5.
It is advisable to add the words - "ginger woman" to the promt. Then you get a more recognizable picture.
evaluation of the model is welcome
Description
Singer
Due to the new site policy, in newly published models, there will be only 2-5 pictures, as an example. so as not to "offend" anyone.
There was something without brains that creates, constantly, new accounts, and put one star in the reviews. of course, without any arguments. not because the model is bad, but because he is an idiot, and he tries to prove it with all his actions.
the training was on a model - CyberRealistic 3.1 in resolution 640x960, and used a dataset of 68 images.
I'm Render this on - EndlessReality v6
if you want to get speed, drive a sports car
if you want to drive off-road, take an SUV
if you want to hammer a nail, take a hammer
it's the same here - if you want to get results, read the recommendations and follow them
I do not recommend using my LoRA with model RealisticVision
personally, I most often perform generation with parameters: DPM ++ 2M Karras, Sampling Steps 36, Denoising strength 0.35, CFG Scale 3.5-8.5
the weight of the model is approximately 0.75-0.85