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://civarchive.com/models/790652/uber-realistic-porn-merge-ponyxl-hybrid-or-xl-and-pony-loras-or-controlnet
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).
Prompting help:
If you notice that it’s not doing what it should do, be extremely light with the negative prompt. And then I re-use the same seed and add more words when needed.
Liability:
In no event shall I or my team be liable for any claim, damages or other liability, whether in an action of contract, tort or otherwise, arising from, out of or in connection with the use of this model. Please render responsibly.
Description
Pruned the model. Reduced the size to 1.6gb from 8gb. Not bad. (Unpruned Version also available (best used for merging))
Rather than using Weighted Sum Merges, I instead merged via Add Difference with the Model A always being used as the tertiary model. This allowed always removing any duplicate data that existed in other merges I was merging with.
By doing the above, it actually fixed problems more frequently in regards to fused limbs, wrong limbs, etc.
Merge Recipe:4 Add Difference merges in total (with Model A always in the Tertiary position as well):
1st Merge:
izumi
sxd-berrymix-merge at 35%2nd Merge:
^above
ZombiMix-v7 at 15%3rd Merge:
^above
3DKX_1.0b at 15%4th Merge:
^above
RealEldenApocalypse_AnalogSexKnoll_4CandyPureSimp+FEET at 25%
FAQ
Comments (46)
Hi Saftle,
firstly, thank you for your models, I'm having great results and will post some of the generated images soon.
Any chance you could explain your how to merge models? I haven't found a clear guide to do so and would love to share stuff with the community too :)
And also, could you add a checkpoint file for your latest version, I'm not able to load safe tensors in invoke and I'm a novice at all that.
Hi there.
I added the ckpt file already. Just use the dropdown next to the primary file :)
In regards to merging, I recommend the following video: https://www.youtube.com/watch?v=xLQcWKI5OLk
Using the X/Y Plots is extremely great at helping you find out which ratio is best. That is also explained in the video.
Good luck on your future merges! :D
Thank you for your help and the link :)
So we still need the old ones ? or we can delete it ?
I would delete the older one. It has slight variations though. So I would only keep the old one if you need to reproduce older renders.
Thanks for information. Yeah I need to trim as much as possible my SD model folder .So any lower size models will be awesome
Rather than using Weighted Sum Merges, I instead merged via Add Difference with the Model A always being used as the tertiary model. This allowed always removing any duplicate data that existed in other merges I was merging with.
By doing the above, it actually fixed problems more frequently in regards to fused limbs, wrong limbs, etc.
I think an add difference merge with C = A is the same as a weighted sum merge.
A + M*(B-A) = A - M*A + M*B = (1-M)*A + M*B
The file sizes are different after doing it for each merge though. The scenario I was trying to fix was when I merged with another model that also had hassan as an example. By comparing it to my original model, it would subtract the hassan portion and then merge whatever ratio I give it.
Merge file sizes can vary for other reasons, even if they're functionally and mathematically identical like A+1*(B-C) vs B+1*(A-C). Which versions of which keys you keep from each model may vary based on which operation you're doing and their ordering, and cause the file size to vary a lot, even though those extra keys don't affect inference and should just be pruned.
fp16 vs fp32 also makes a big difference in what a model outputs (not necessarily worse or qualitatively different on average, but different nonetheless), which based on model file sizes I'm guessing is a difference between v1 and v1.1.
You could be right that add difference vs weighted sum merges that are naively identical are functionally different, though in the few cases where I've tried comparing outputs from the two I've found them to be the same. You can really only find out whether this is the case by doing the two merge sequences at the same precision, ideally fp32, and comparing a grid of output from each with the same prompt, seeds, and selected VAE (in order to override the built-in VAE, which is one of those keys that can be different depending on which model was A/B/C in which merge step).
Your math is ok but I don't think it apply here because if M = 1, one would get only model B with your equation whereas in reality you would normally get A + part of B that is not in A
But actually Ijust tried with 2 dreambooth models and it is exaclyas you say. If you use M = 1, it keeps only model B....
I think we need to find other ways to merge
Hi Saftle!
This might be interesting to you.
So, originally I downloaded uberRealisticPornMer_urpMv1.ckpt (8gb) and used it as a base for a merge (add difference). All went fine, the B was perfectly visible in A (your 8gb model). I used the float16 so the result ended up being 6GB.
Then you made a 1.7GB safetensor version and I thought great, I can remake the model and save some space.
What happened is when using AUTOMATIC1111 and merging my ckpt model into your safefensor model I've got some "potatos".
I saved several prompts to recheck them on new merge and expected them to be similar or the same. But what I got was picture of a rock, picture of a city map, a picture of some pattern. Seems like the model collapsed and completely ignored the prompts.
So what I did next was to change your safetensors model into ckpt model (was then 2gb) and it looked identical.
Then I merged my ckpt model into your model (converted to ckpt model) and then the subject from my model was present in yours (success, right? well kinda).
Not only the same seeds gave different results, the whole effect was kinda horrorish. The person was recognizable but the whole composition was quite surreal.
This is something you perhaps want to investigate because at least for future releases you may still want to release the full precision and the pruned versions.
But overall - great work, I will still be using the original model, the size as they say, does not matter :)
Hi there. Thanks for the write-up. I was told the same thing recently, which is why we should always merge with the non-pruned version. I also include the non-pruned version for those that do merges. Click on the arrow on the right. I'll always release a non-pruned version for every version I release.
Glad you like the model though :)
Yeah, the model is amazing and it works fine with embeddings :-)
I will make a review with image soon but I need to generate some nice non custom-merge content :)
I'm glad I'm not the only one reporting this, it's actually a great feedback that merging should be done with non-pruned versions! :)
Cheers!
Hi Saftle,
That URPM can use with the inpaint mode?
It can, but it's not meant for it. I learned how to make an inpainting model recently. I'll go ahead and make one in the next release :)
How can I use this model with DreamBooth? I always get an error when converting to diffusers
There appears to be a bug of some kind on Dreambooth's side, since all I did was merge via Automatic1111, in regards to the unpruned model. Here is a similar issue on Fast Stable Diffusion: https://github.com/TheLastBen/fast-stable-diffusion/issues/1257#
It doesn't work with NMKD GUI, I've tried everything but it doesn't work :(
Did you try the non-pruned model? You can press the arrow down to see all downloads. If that doesn't work, I would create a bug report on that UI, since all I did was merge models via Automatic1111 in regards to the non-pruned model.
The VAE is now included in the model. It is no longer required starting with version 1.2. Hopefully this solves your issue.
it works now (with the 1.2v) ^^ just convert the .safetensors file back to .ckpt in the NMKD GUI under "Developer Tools", place it in your "models" folder and select "None" under VAEs.
I Mixed this Model with Analog Diffusion and got Amazing Results. Could i Contribute the Safetensors File somewhere?
Interesting. I briefly tried to mix it, but I didn't get good results if I recall correctly. Or perhaps it hurt the porn aspect too much to include it, which is the case for alot of merges. It's extremely hard to get good visuals, while not reducing how accurate the porn scenes are.
However, I just finished with URPMv1.2 where I changed alot of the current recipe which may play better with Analog Diffusion.
Would you mind throwing me the ratio with which you merged at? I could try and incorporate it, if it works well :)
to my inbox :)
Downloaded the latest version and now URPM generates only black images ? All my other models work perfectly fine. Weird ?
Any idea please.
Did you load the required VAE? Installation steps are in the description.
It is also possible that you have the NSFW filter enabled accidentally that Automatic1111 comes with. It makes everything NSFW black.
I updated the automatic UI and immediately afterwards the black images started appearing. I tried all 3 URPM versions and they all do it, while my other models don't. New update is breaking something with URPM.
@btaco6, are you using the pruned version or full model?
I have been using pruned versions so i downloaded the full 6gb which seems to have fixed it. I am able to even go back to older versions and the black box is gone for those too.
Yeah the pruned version is a little broken in 1.1. I'm uploading V1.2 which fixes it. There should be no differences between the 1.2 pruned or unpruned model aside from useless junk removed.
Thanks saftle. Dunno what happened for a few hours with my Automatic1111. I had issues with all models generating nonsensical and truly horrible images. Super weird.
I posted about that on Reddit and actually quiet a lot of people had the same issue. It lasted for a few hours and finally a git pull downloaded an update that fixed whatever caused that.
So it wasn't your model ;)
Cheers.
having a hard time adding the file, I put it in the right place, add that line to the quicksettings, but when I try and select it in the new drop down, it quickly jumps back to 'auto'. any tips?
Luckily the VAE is now included. So if you grab the newest version (1.2), you can ignore all of that previous installation regarding the external VAE.
Where would I find the VAE?
It's the 2nd link in the description. Follow the instructions in the description further down after clicking "Show More" :)
I should've clarified. I meant how do install it for InvokeAI?
It is now included in the model. It is no longer required starting with version 1.2 :)
How do install the VAE for InvokeAI?
In the newest release of InvokeAI, you should have a model manager icon in the upper right hand corner. You just edit the model in the list and add the local path to the VAE.
However, I'm in the process of uploading V1.2 which includes the VAE into the file itself, removing the additional requirement, so you can also wait for that, if you'd like :)
The VAE is now included in the model. It is no longer required starting with version 1.2. Hopefully this solves your issue.
I actually installed the wrong model type in your list. I have it working great now. Thank you for your help though. Have a pleasant day.
What did you do to get this working in Invoke? I cannot seem to get the model to work using Invoke at all. Just produces weird results that do not match prompts at all.
I was having the same issue with it @stablydiffused I had the same issue, but I used the pruned version and it worked fine
I did end up switching to the pruned version as well. That one works fine in InvokeAI.