Do you want to support me? Consider being a patron on Patreon to vote on train polls and access stuff early. ❤️
MS Real is a new collection I'll be doing. Instead of Anime images like I do normally in "Murky's" MS Real will be focussed on realistic models. So far as I tested it works on anime models too.
Trained with 100 images and 768x768.
Tags you might need are
solo, ass, pussy, lying, anus, on side, looking at viewer, nude, spread ass, ass focus, presenting, ass grab, spread anus grabbing own ass, hands on own ass,
Use the LoRA weight around 0.6 and 0.8 for the best results
Known Problems
-Like all the realistic models it will create body horror time to time.
-Too close on ass. Dataset was mostly close up shots so It can get really close time to time. Try with lower weight
Discord server feel free to drop and say hi ^^
Description
Trained with lower lr and longer epoch.
Working with urpm chillout and anime models
FAQ
Comments (12)
you should make an ass licking lora, we have yet to get a good one and i think it would do well,
95% of the images this lora makes are complete mangled mutations. And I never managed to get a single decent image at 768/768, which is odd since that what this model was trained on. At 640/640 it manages to do better, but still you have to go through a lot of seeds to find something that's not terrible.
This was using URPM, using the exact prompt given for the sample images to have the best chance for success.
dosent matter what the LoRA was trained on. You always get better consistency with 512 if the model you are using was trained on 512x512 images which most models are.
As beric said. Just because It's trained on 768 doesnt mean images should be generated at 768.
Most of current models is better at 512. Use 512x512 and use hires fix to upscale it.
@Murky I assumed 768x768 would work since the example images were all in this resolution. I rarely have any luck with hires fix, it produces mangled images or worse images or it doesn't actually add any detail so that the larger image is just an expanded version of the smaller images. Or it removes all the texture from the skin which is actually worse. And it takes so long to make a picture with hires fix, and by the time I have tried and failed repeatedly to get something good I've wasted 10 minutes on it.
@Icanhazusername 1.4 and 1.5 models works best with 512. if you check like original stable diffussion checkpoints they actually have seperate ones for 512 and 768 ones.
Training on 768 is helpful to give more detail when you use hires fix. you can still use straight 512x512 too
@Icanhazusername Yeah you're doing it wrong. get a upscaler from https://upscale.wiki/wiki/Model_Database i suggest ultrasharp, put it inside stable-diffusion-webui\models\ESRGAN to make it available in hires fix and use 0.30 denoise and below. If you think hires fix takes too long you either dont care enough about the images you're generating or it's time to upgrade your GPU. i hires fix to about 1200x1800 and it only takes 30 seconds per image with a 3070 ti.
The ass is too fat in this lora
coward
isn't that a good thing?