Heh, in all honesty creators set what they feel is nesscary to recoup costs, I'm aware it's a struggle bus, but Saftle deserves a lot of love for all the hard work he's done (Look i'm sadly probably not going to use it much because i'm not a porn user XD -- but I think saftle's a mazing and he's an FF nerd)

@saftle https://civitai.com/posts/7533751 see? works with pony lora already, it's cool.

@saftle i know right? i thought it would create a blurry pixelated mess, but it's surprisingly good.

sweet home alabama riffs ... xD

If my sister was as good looking I'd French kiss her too.

@sjukfan Of course there's the slight difference between a step sister and a sister, but what do I know... Din sjuke fan xD

we still need flux thou

Doesn't merging this model with realvis work? Just wondering if you tried and if it's possible! :)

I haven't tried mixing the ponyRealism model yet, but using the URPM model as a base model and other models as refine models (at 0.4) hasn't been a problem under Fooocus. However, I will be interested to see a person Lora training under URPM, which I hope to try this week.

@mykeehu I'm pretty curious about that too! :)

I'm training Lora now and running her for the second time, but it still be too strong.

I always do cosine training on Prodigy and control the strength of the training with the d_coef value. Now it's set to 0.2, with normalization, but it's still very strong (dim/alpha: 32/32) It was good on SDXL and Pony models with 0.3 d_coef. So either this model learns very fast, or I have to increase the dim/alpha (if I halve it, it learns even faster).

So I keep experimenting.

Update: I also stopped the third training, I simply slowed down the training to the minimum, but I still couldn't make a good Lora for this model, I'm sorry :( It would have been really good.

But that doesn't really justify your complaint either, considering in 2 weeks it will be public anyway, if you don't want to spend 2k buzz just wait for it to be released for free then. :) Just turn on notification for when it does. ^^ also URPM was a gamechanger for 1.5. And now they actually managed to merge xl and pony with their cybermagic its a gamechanger for this gen too. :P

@saftle you are a saint. this didnt need any explanation. the only thing you have to tell us is was the ceo/money investing guy andrew tate yay or Nay? I've had the same idea, starting a model agency with this? but use cases are still rare and hard! keep it up!

@saftle I payed the buzz just to show my support! Please don't ever quit! <3 And I'm also wondering, did adding the "URPM dataset" to SDXL add any new vectors/tokens or did they edit/overwrite existing ones? Hope my question makes sense at all. :P I used to get amazing results with the 1.5 version and I'm hoping to get the same type of behaviour from this one. I have some more questions about the dataset though, but to clarify I have no interest in the data/images itself but instead wonder about the type of language used. Like does the dataset use a form of booru tagging or natural language? Was it interpreted with gpt, blip or something else entirely? And if natural language was used were there alot of komma's used to seperate sentences? So basically Im wondering if besides the language used of pony and xl there is also added stuff from the urpm data. Because I noticed that I always get the best results by prompting in the same manner as the dataset was captioned. :P The answer on these questions would seriously save me many hours! <3

@saftle Wow, thank you so much for answering so quickly! I really appreciate the amount of effort you put into this. I'm about to do my first run with this killermodel. I'll always do a multitude of tests which usually take a few days of extensive prompting. But I'll make sure to report my findings to you! <3

how to work lustify properly ?

no, everything generated was raw, no vae used, i know cause i downloaded his above images and used png info on a1111. https://civitai.com/images/33587896

@saftle I'll see if i can find it

@saftle https://civitai.com/posts/7915987

i dont think furry will work with a realistic model, furry is limited to cartoonish/anime stuff, coz only these have fury data.

@KawaiKillerXoXo There are realistic furry models; furry content isn't necessarily cartoonish or anime. One could consider photorealistic furry a different artform from the digital painting/cartoony/simple stuff. Photorealistic within the bounds of SD, that is. There's other models that do furry content much better, or at least, easier. If furry is something one does frequently, look for a good furry model, not a sex model like this. My humble advice, anyway.

@saftle will do, and thank you in advance

if his pateron is down, thats sketchy.

are you using score_6 or score_6_up by any chance in the prompt?

@KawaiKillerXoXo 9, 8up, 7up in positive; 4, 5, 6 in negative. Let me know if you have any insight, hugely appreciated!

I'm also having a problem with lighting for the PonyXL Hybrid.

Had one batch turn out absolutely perfect, if not a bit bright, then everything after that with near-identical prompts and LoRAs all came out completely unrecognizable, with the most extreme blown-out contrast imaginable (just bright whites and dark blacks with nothing in between).

It's a real shame, since this model is obviously capable of producing great results. Just not exactly sure what's causing the issue.

@saftle just gave it a try and the image generated using baked in VAE vs. the one you linked were identical, appreciate the response all the same - let me know if there's anything else you'd like me to try. I bad mouthed your 1.5 model the other day so I wouldn't mind finding a way to make up for it lol

I've figured out the over exposure thing happens with too many Lora's. Cut out any that aren't essential and you should be good.

@saftle yes, also experiencing the over exposure, high contrast issue even with just one LoRA applied. If I reduce the strength to 0.6 or something, it goes away, but it's very temperamental. Strange.

@saftle Sorry, haven't had time to replicate it. Also, if I recall, it is hit or miss (which makes it frustrating to generate since random seeds produce the problem. I believe I was getting it just Youthful Pony and at various strengths above 0.6.

@saftle np! I love the hybrid model, it outperforms other NSFW SDXL and Pony models. Can't wait for a v2!

Check my uploads, About 7/10 succesrate with hands. The helpful part is: "perfect human hands fingernails", jewelry ring.
And you can just put "bad hands" in negative. Just the tag, not the TextualInversion/embedding.

I recommend you check my gallery if you are having trouble with anatomy and read my journal Ultimate guide to heterosexual NSFW SD1.5 (unfinished-WIP) | Civitai! Which is still mostly accurate.

And for URPM-Pony I recommend a CFG of 5. Any higher increases the chances of getting overlit scenes. And putting spotlight in negative somewhat helps too.

Happy prompting! :)

@saftle Thank you for your response. I hope you can make it more 'generalized', because it does make nice images, I just feel I'm fighting it. But just for the nice response, I'll give you a bit of buzz.

try: "pale translucent white glowing ghost skin" in negative prompt., this helps me for some of the generations! There is a tag that is causing it somewhere. Also on high CFG it gets a little worse, try a lower CFG like around 4-5. The model still has great responsiveness for CFG this low. Anyway, I am honing in on the problem. Will come back to share once I found it!

any news? for now I add a dark lora to compensate but that makes the background too dark :/

@bewunderteradler112 adding the tag "paper" seems to work for me as well. Still experimenting with different tags to get the succesrates higher though. But slowly getting there but been a bit busy.

And "dark" also seems to make the images a lot less saturated. Will come back here if I find some more.

@c666d thanks for the feedback!

@bewunderteradler112 Also try in negative: "spotlight, amateur"! These work wonders for me! :O
Alright, will come back for another report soon. :P

img2img from a black image (or your desired gradient...) at a high denoising and not only does this problem go away but you have much more control over the lighting in your output

Check my profile for tan Slider, this will help you out

Use another checkpoint? Use lora?

@OkiMoki667 I can try the Lora and see how it goes.

heck yea nakked womens!

People use to make fun of me for being a digital hoarder. =p

@mike_frett Hopefully you reupload some of the lost data to Tensor Art or SeaArt, they're not run by Communists

@saftle Celebrity LoRas are gone. Actually just hidden. A lot of censorship is happening to please the credit card gods...

@mike_frett You might still find them on civitarchive

@Yukihayashi the ones that delete loras so they are monetizeable are "communists", and the ones who freely share are not? make it make sense 😂

Civitai's policy changes

It seems that Civitai removed real-person likeness content in May 2025, likely due to pressure from payment processors like Mastercard and Visa, as well as new legislation.

How is it a pony?

Put in \StabilityMatrix\Models\StableDiffusion\