Uber Realistic Merge (URPM) - Alphas [LEGACY Versions]

Daily Release Campaign has ended!

See here for more info: URPM is back - Daily Releases incoming!

💖Join the Patreon if you want to support our work & even earlier access to models being beta or alpha tested!

🗣️Join the URPM Discord for Updates or Support!
 ---
My SDXL and Pony-SDXL Hybrid models are now available here: https://civitai.com/models/790652/uber-realistic--merge-ponyxl-hybrid-or-xl-and-pony-loras-or-controlnet

About this version:

Over the last few months I've busy working with a trained model from a friend and trying to get it working with URPM. However after many attempts, the merges are good at (and even accidental ), but worse in other ways, so it isn't a great upgrade to URPM itself yet. It can also fall apart sometimes if you try to describe the people in specific ways.

So until I get URPM with to a stable place (that doesn't hurt other aspects of the model), here are all three alpha builds for .

CE1-ALPHA has the most consistent results

CE3-ALPHA has the best skin

CE2-ALPHA is somewhere inbetween.

I'd try all three.

Remember these are ALPHA versions! Your mileage may vary. Please temper your expectations. I hope you all enjoy experimenting with these builds!

Thanks to my anonymous friend/contributor for providing the model to merge with.

Example Images:

None of the sample images were altered, upscaled, etc. All can be reproduced using the metadata via the unpruned model (however the pruned model should now have the same results).

Xformers: Off
ETA Seed Noise Delta: 0
ClipSkip: 1

Prompting help:

Female on Female (usually every render is decent)

woman another woman's , ((())), ((detailed )), ((detailed faces)), 2girls

negative: ((blurry)), duplicate, deformed, makeup, cartoon, animated, render, child, childish

Male on Female Cunnlingus (might take 4 tries to get a decent render):

man woman's , ((())), ((detailed )), ((detailed faces)), 1man 1girl

neg: ((blurry)), duplicate, deformed, makeup, cartoon, animated, render, child, childish

If you notice that it’s not doing what it should do, be extremely light with the negative prompt. Example: ((blurry)), animated, cartoon, duplicate, child, childish

And then I re-use the same seed and add more words when needed.

Just like any merge that contains merges with Stable Diffusion 1.5, it is important to use negatives to avoid combining people of all ages with . This is sadly unavoidable without adding negative prompts, until there is an embedding or the like that can help automate this process. Here are a few things that I generally do to avoid such imagery (and will start representing this change in future version examples):

I avoid using the term "girl" or "boy" in the positive prompt and instead opt for "woman" or "man". The only exception is when I am specifying how many women should appear in a scene, which I would then use "1girl" or "2girls".

In the negative prompt I use: child, childish.

This has helped me prevent any kind of accidental imagery. I know a lot of us are used to using the term "girl" for "women", but AI can't understand the difference.

Merge Recipe:

As an open source advocate it is important for me to help provide the "source code" with everything I do. You can find the merge recipes to the right of every version below.