It's funny that there's almost always similar comments (about being better and more consistent) for every new attempt for a photo-merge. I haven't tried this model, so maybe you're right and this one really works, but after so much time I believe pony v6 is not compatible with photo realistic stuff. As I understand it was trained mostly on cartoon-ish images. And all merges so far, including the one you mentioned, gives only some 2.5D vibes, no realism there. Not a bad thing, btw, a decent merge with such models can give quite unique (still illustrative) output and with the help of PAG some extra details. I think we all want to transfer pony's ability to follow the prompt to some photo-realistic models, maybe v7 will give such an option (I'm wondering if they wait for SD3 or things are already in motion).

@green_anger I read those comments as well and typically gave them a try haha. This one passed my initial testing at least.

I don't think photorealism is genuinely possible either. But, there's at least a lot of attempts and some better than others. I found most of the other models would do portrait shots great but the faces just fell apart for anything else; and the style would fall away the farther you got from portrait shots as well.

I'm not gonna lie and tell you this is juggernaut with pony prompts (it's not), but, it might end up replacing everclear for me at the very least. I put up a sample picture to give an idea, as I wanted to see how it worked on non-booru tags as well. I was able to get rather cool scenic and diverse images which always felt hard from pony models.

Pony is waiting for SD3 for sure and preparing for it. They've got some notes for that on discord. I think that'll be much better and he explicitly mentioned at least having better training and captioning to make photorealism easier to achieve even if it isn't built in directly to v7.

@green_anger FWIW specifying age and ethnicity goes a long way, every example has a weird 2.5d doll face like you're saying but I'm pretty happy with how my generations turned out when I dialed in details about the subject.

@Talnik Alright, I need to try it then, from the examples below (some of them) I can see your comparison to other models is quite accurate, it does look more realistic than other models.

Regarding faces, distant faces are almost always ruined, it's pretty standard for the whole SD series, including XL. What I find weird for real-pony attempted models is horrible eyes. When the face is about 1/3 - 1/2 of the picture, then it's fine. The farther it gets, the bigger chance to get some red/horror eyes. Standard XL or non-real-pony don't suffer from it, while real-pony merges very often produce something bad. And yes, we can always inpaint (I find diff-diff can work miracles more often than not), but I don't even want to touch those eyes when I see them, the only wish is to print it and burn. This model also shows similar tendency, half of the listed examples including yours have it, but it's not as severe as I described. One more reason to agree with your evaluation.

[--Some lyric about SD3, can be safely ignored--] I've seen some early preview of SD3 from a guy who supposedly works for stability ai now and have access, and it wasn't impressive. Yes, it's not a final product and I remember the original base XL quality (hey, refiner, where are you, buddy? - I'm wondering if most people even remember what it is), so it's going to be better after a couple of months and applied community enthusiasm. My biggest concern with SD3 is requirements. I don't think they've published it, but all GPUs are being bought out by big corps, it's getting harder to generate locally with the latest tools that get updated every year now (SDXL was published last summer). So, if they increase requirements, it may get a blocker for many people including myself (8GB VRAM). On the other hand, they say there'll be a range of models with different number of parameters, so maybe things will not be that gloomy. Hopefully all that range won't require separate training and can be somehow compatible. Everyone's getting into pre-SD3-waiting mode, let's see what it'll bring.

@malias693 Age is hard to control even in standard XL, in my view it's even harder in pony. Looks like it wasn't the focus during initial training (of SDXL) and it's ok, you cannot include everything and people are not the main focus, only part of data. In my experience there's a few age categories that can be somewhat controlled: (1) kids and teenagers, (2) mature up to ~30yo, then there's a huge gap and (3) old people. And if you look closely at many examples (not specifically here), even when it says something like "young woman" or "18yo girl" the model gives you something closer to what was put into negative (not going to get into this, I think you should get what I mean here).

Ethnicity really varies based on fine-tune material that was used, but I don't play with it too much, so not sure.

@green_anger good points, I moreso mean putting "child" into the negative so you don't end up with baby faces. I couldn't get this model to produce anyone older than their 50s.

@green_anger caveat which @malias693 goes into would be this checkpoint is moderately young biased, but overall that's not the hardest thing in the world to control. It's not perfect but it definitely feels like a good next step in pony photorealistic, at least until SD3 comes along!

Interesting, I'll need to tinker with this one a bit more. Had bad luck trying a promt used on other Pony realistic models. Was about 1 out of 8 successful rate on key aspects of the prompt.

Thumbs down and laughing but yesterday the gallery was filled with underage looking gens, and now they've all been deleted.

@VelexiaOmbra  Odd, don't seem to remember that... Are you sure? maybe some other model or website that you are browsing?