Anya / Merilyn Sakova fakes

[Upscaler]

I trained an AI model using 100 photos of Merilyn just so I can make these.

 

[Upscaler]

Are you sharing your custom models to any site at all for sharing? This model and any future model you cook up would be awesome to use.

I didn't want to bore people with technical details but I actually trained a hypernetwork which as I understand it is more like an elaborate filter that twists a model to look like the photos you trained it with. That makes it quite useful as you're not tied to any one particular model and can use it with a range of them, including the best, latest models as they come out. I don't have a problem sharing mine but unless you know how to use them with stable diffusion it won't do you much good:
https://www.dosyaupload.com/48Lxb/hyper_merilyn-2200-10000.pt

woah, that's cool! She looks like the sorceress from the video game 'Dragon's Crown' in one of those pictures.

Here's some more I made of her as a witch. The facial resemblence isn't as strong as the first picture.

 

[Upscaler]

you could quite litrally make her do what you want...!!

Eh, kind of. I'm still limited by the models and they are notoriously bad at certain things. Especially X-rated stuff, I haven't found a model yet that produces that kind of content with any degree of accuracy or consistancy.

It works a little better just painting over existing porn with her image but even for that to work you have to find a starting model with roughly Merilyn's proportions and there's not that many out there!

 

[test902]

well I appreciate AI art in general. If you don't mind my comment, MS in AI art looks like a younger and even bustier version of the actual MS in her (or her breasts') prime.

 

[speller007]

Eh, kind of. I'm still limited by the models and they are notoriously bad at certain things. Especially X-rated stuff, I haven't found a model yet that produces that kind of content with any degree of accuracy or consistancy.

It works a little better just painting over existing porn with her image but even for that to work you have to find a starting model with roughly Merilyn's proportions and there's not that many out there!

still, the effort involved.. this could become quite comical if you included animals she still looks extremely realistic

I was thinking milena velba in her prime or bettie ballhaus before she got her reduction

 

[bombasa_juan]

i spended my whole afternoon trying to set up the automatic1111 stable difussion and couldn make it, so glad that you keep posting this hahahah

 

[vveeaaee]

can you share more about the tools you used?

 

[Upscaler]

can you share more about the tools you used?

98% of it is Stable Diffusion and the other 2% is using Photoshop to erase some unwanted elements from the photos I used to train the model.

 

[celsior360]

Eh, kind of. I'm still limited by the models and they are notoriously bad at certain things. Especially X-rated stuff, I haven't found a model yet that produces that kind of content with any degree of accuracy or consistancy.

It works a little better just painting over existing porn with her image but even for that to work you have to find a starting model with roughly Merilyn's proportions and there's not that many out there!

Yulia Nova or Lucie Wild will be a good candidate

Is there a whole link where we can download all these wonderful images?? Thx in advance

Is it possible to make a small clip with the AI technology??

 

[joehiffle]

I trained an AI model using 100 photos of Merilyn just so I can make these.

Marvelous work! Looks great!!

 

[Lucaxyz3]

Is there an app on android that allows the AI image training?

 

[cbanner]

I really love those Merilyn ai images. Do you guys have made more of them ?

 

[fakeman2002]

This is definitely a fake creation that I made. Just imagine what could have happened next.

 

[dubh3d]

They are excellent. I'm glad to see someone using AI to expand the fantasy with hardcore.

 

[Upscaler]

It's been 10 months since I made those original photos and the technology has progressed. Nowadays I'd use a Lora instead of a hypernetwork and would train on the SDXL base model instead of the increasingly obsolete 1.5 base. Development on X-Rated or even NSFW SDXL models have been much slower than I anticipated however so I just haven't done any custom training at all.

 

[anriokita]

It's been 10 months since I made those original photos and the technology has progressed. Nowadays I'd use a Lora instead of a hypernetwork and would train on the SDXL base model instead of the increasingly obsolete 1.5 base. Development on X-Rated or even NSFW SDXL models have been much slower than I anticipated however so I just haven't done any custom training at all.

Whats the slow up(with SDXL) if you don't mind me asking? I ask because I want to be able to do this myself once I get a better PC.

 

[Upscaler]

Whats the slow up(with SDXL) if you don't mind me asking? I ask because I want to be able to do this myself once I get a better PC.

Probably a combination of inertia (people sticking with SD1.5) and computing power. SD1.5 was trained on 512x512 images while SDXL was trained on 1024x1024 images (there's also some architecture differences but just ignore that for now). Doubling the size may not sound like that big of deal but training with bigger images requires more vram which in turn requires better and more expensive graphics cards.

I've only ever trained models on one particular female subject. I'd like to make my own general NSFW model that at least had accurate genitalia but I've never done that before and I have no idea where to start.

 

[WatsonBlue]

It's been 10 months since I made those original photos and the technology has progressed. Nowadays I'd use a Lora instead of a hypernetwork and would train on the SDXL base model instead of the increasingly obsolete 1.5 base. Development on X-Rated or even NSFW SDXL models have been much slower than I anticipated however so I just haven't done any custom training at all.

It's been 10 months since they were created? Sounds like you're a bit rusty and probably could use some practice creating new ones to get up to speed with the newer processes.

 

[Upscaler]

It's been 10 months since they were created? Sounds like you're a bit rusty and probably could use some practice creating new ones to get up to speed with the newer processes.

If only it were a straightforward process. It took me a half day to put together a new data set and hours to train a range of loras. The problem is deciding which one to use. You can go for flexibility and have a lora capable of multiple art styles but the accuracy suffers. Or you can go for the highest photorealistic accuracy possible but then your lora will only able to do photos. Finding the right balance between those 2 extremes is the hardest part and requires a lot of testing.

 

[WatsonBlue]

If only it were a straightforward process. It took me a half day to put together a new data set and hours to train a range of loras. The problem is deciding which one to use. You can go for flexibility and have a lora capable of multiple art styles but the accuracy suffers. Or you can go for the highest photorealistic accuracy possible but then your lora will only able to do photos. Finding the right balance between those 2 extremes is the hardest part and requires a lot of testing.

No disrespect meant, it was a bit of a tongue-in-cheek gentle request for some more of your great work, and I thank you for putting in the time to share more of it. It really is awesome.