This lora was trained on a curated dataset of 16k+ images to give realistic styled images a more "social media post" style. It works on SDXL realism models. I tested it on Lustify as its the best realistic model I can find for SDXL these days, but with many NSFW tags in the negative.
If anyone knows a better "generic" realism SDXL model let me know!
Description
FAQ
Comments (2)
16K pics really? Thats huge, its enough for a trained checkpoint. Im sad, because on the dmd model, what i use, it looks oversaturated. Maybe just on dmd2 model, i dont know.
You should make a model from this. Best general models are still Juggernaut and Cyberrealistic. Maybe Epicrealism.
I've never tried making a checkpoint before but I'm definitely willing to give it a go. From my understanding I thought you needed ~million images for a checkpoint finetune. Also I use dmd2 with lustify workflow and get good results. I have dmd2 turned down a bit though, I think something like .75 on model, and .55 on clip maybe (on my mac, away from my pc for a while so I can't check exactly). I think for steps I have around 10, and cfg is pretty low, 0.9-1.2. I have the upscaler set to 40% and denoise to like ~25%. Also definitely play around with the tuning of the lora settings as well. I find that mixing it with some of my loras that have small subsets of some of the same dataset (200 images) works well (although I've never released them, but maybe if there's feedback that people would want it?)
If that doesn't work I can take a look and add another reply once I'm back to my home machine!