RosesForMen V4.3:
All the credits are on the "about this version" section.
I think this version is more solid than the others, the improvement is very noticibly.
As always, is not perfect, but this time the results got better, and I didn't have major issues with this model in comparison, it needs some fixes on fingers and hands, maybe an eye but overall gives good results, and you can use this same model for inpainting simple things.
Pros:
-Better Faces.
-More stability.
-Better quality.
-Cocks looks more realistic. (not perfect)
-I didn't even use negative prompt.
Cons:
-It seems to have issues to produce "far shots", I still don't know how to manage that.
-It doesn't understand "Flaccid" or "erect" cocks anymore, but can produce Flaccid without asking for it.
-Still needs upscale to get the best results.
Another tip: It seems to fix some issues by doing 15 steps instead of 10, but sometimes 10 steps follow the prompt better so it depends on your preference.
My recommendations for better results:
Sampler: DPM++ 2M Karras (it also works well with Euler a / DPM++ 2S a Karras)
Upscaling method: Latent
Denoising strenght: 0,7
Upscale by: 1,3 to 1,5
Steps: 10 to 20
It's not perfect and also in some images may have some gray spots like in the examples, and since I have almost no knowledge about merging and how to fix some things or maybe make better prompts, I'll let anyone use this merge to improve it and give feedback to make it even better.
First I want to thank and give credits to all of these amazing and talented creators for sharing their models to everyone, so the models I used in this meged base models are:
LoRas I used were:
Again, thanks to all of these creators, and also as they request, is not allowed for commercial porpuses.
The configuration I used:
Sampling method: DPM++ 2M SDE Karras (I haven't tried another method so feel free to try)
Sampling steps: 20-30 steps
Resolution; 768x1024, 1024x1024
Some words that worked for me in the prompt are "photography, ethereal, aestheticism, cinematic photography"
And for negative I used "((head out of frame)), asymmetric face, blurry, Facial hair, anime, cartoon, graphic, text, painting, crayon, graphite, abstract, glitch, deformed, mutated, ugly, disfigured, head out of focus, head blurry, necklace, big forehead, earrings,".
But you can try yours too.
Since my pc is low specs feel free to try it with other methods, different steps and everything you want, I apologize for not testing it more.
I will be glad if you use and share some pictures of your own results.
Description
After a lot of merges and some training by myself, I think I could improve this model, and I personally love the results.
It has better undestanding of the prompt, better sharp images, but as any model is unstable, so of course sometimes gives bad proportions and also the gray spots still there (I know that can be fixed by bakin VAE, but I didn't like the results with VAE).
It does need some inpainting on faces, so I did use inpaint on some faces and to fix some of those gray spots in some images, just to let you see the potential of this model. (nothing else on the images modified)
You have the prompt on every image so you can copy my config, but feel free to try anything you like.
Tip: Upscaling the images seem to fix some of the visual problems.
Credits for some models I used in this merge:
And I apologize if I forgot about other model cause this one of many merges I tested...