Does everything a PonyXL model can, but "nearly" photorealistic!

@Zovya Thanx, i found the embedding but no v2 Version. So i have to wait.

@Zovya I am using a lora of my own for that type of skin, and this version "neutralizes" it by a lot, compared to v2

The description on the model says "So I've purposely kept it from getting too photorealistic for the moment due to uncanny valley issues." So no, this probably isn't going to be the model you want. Personally I hope they keep making it the way they see fit.

I've used a lot of the photorealistic ones and there's still a lot of quality issues with all of them. The compromise this one makes is solid.

@dfjunkmail Just stating my impressions for this version, and looking forward for the next one which I hope will be more photorealistic (at least for me)

@SAC020 it's been tough finding the perfect model for flexibility and realism. the previous mklan pony model was pretty good but the new one this week, although the most realistic and least glitchy one i've used, takes a bit too much coercing to get the prompts right.

I've been experimenting with this model for more realism this morning though and think these are my favorite settings so far. Still more CGI-like as expected but provides a little less "plastic" imo:

Steps: 25, Sampler: DPM++ 2M SDE, Schedule type: Karras, CFG scale: 7, Size: 640x768, Clip skip: 2, Hires upscale: 1.5, Denoising strength: 0.7, Hires steps: 18, Hires upscaler: 1x-ITF-SkinDiffDetail-Lite-v1

are you using his negative/positive embeddings? I've had issues with the ZPDXLrl (positive) embedding & celebrity lora's. (can lead to garbled mess).

@SleezeBagDiffusion No, they also change the appearance a bit too much. Of course, most loras will have that effect too. But I try to minimize the effects as much as possible.

@d_A_v_I_d  I see what you're saying now.. celebrity LORA's especially. I still think Fast Pony Merge 7.5 specifically (not 9!) is best at handling Razz's celebrity Lora's. Something happened in version 9 that isn't quite as good. I'd suggest trying it if you haven't already!

popular not meant best

ooooh, that's really too bad. I loooove Zovya, I'm really hoping things will be gender balanced though

I don't even mind if it can generate more female images, it is more of accuracy and cost, now Civitai online gen cost money and no matter how many times I put in male, man, 1man etc the result sometimes give more female. I am liking the realistic look that some of the results come out though.

I like to add a lot of pussy juice, spit, drooling, but it only goes so far.

Various sex positions with any lora, even one with a SFW training set. The best I've seen outside of pony is Juggernaut and it completely fails at NSFW with a SFW trained lora.

It's understanding of NSFW concepts is unparalleled compared to any other models I've found, it also knows lots of obscure anime characters

ah ok thanks

@amazingbeauty also general booru comprehension. Due to some sort of overtraining pony is A LOT more specific.

@Volnovik now i have one new question, what is 'booru comprehension' ? i have no idea what that mean

@amazingbeauty I mean prompt adherence.

@amazingbeauty Specifically, the kinds of prompts you would see on an imageboard site like rule34. Whenever I have trouble getting a prompt to work, I go to rule34 and see if I can find a tag phrased like what I want and try that (e.g. "breast_grab" or "ass shot")

A follow up on this: it seems ((female ejaculation)) works sometimes. Not quite as well as the images I posted but better than what I was getting with prompts involving squirt.

This is pony. Try names from booru etc.

@Volnovik oh, you mean I should use anime charaters, hoping they look good on this photorealistic checkpoint?

@settima_ai this is a pony finetune. Thats it. Kelly Brooke is not present there (I guess) for better or worse. And it is weird to expect her here. But you can also search for booru tags describing facial features, it can give an interesting result.

@Volnovik where can I find a list?

@settima_ai any booru site like safebooru, danbooru, gelbooru, aibooru etc. Look for tag groups. Also you can use Ranbooru extension for randomising stuff (not for this checkpoint imo). And yeah photorealism is not the feature of this exact model, at least yet. Also I advise you to check Pony and it's authors threads just to understand what it is, why it is and how funny the whole thing of it becoming best anime model so far (in prompt adherence) is.

Same problem. To get different face i forced to completely change the prompt

Use a face swapping website.

@TubbyCakes oh, totally forgot ipadapter with face transfer

@Volnovik Sorry, I cant find danbooru tags for "face geometry", just expressions. Even if, I'm afraid it wont enought to fix a face to a precise result. For example, even with "oval face & big eyes" there are infinte different faces...

I believe it will need to be in the same way as for Pony, but I'm still figuring it out myself.

Concept art darkside lora by https://civitai.com/user/prgfrg23 does it for me. But it looks like he deleted that pack from civitai.

Edit: He did not, but I cannot find it via search for whatever reason: https://civitai.com/models/264290?modelVersionId=446346

Thank you @Volnovik, civitais search function is what it is.

@SlimeGirl I mean they can theoretically, but as they have been trained on other models the quality will drastically go down usually.

@Maya_H_123 makes sense, thanks. it also explains why I’ve been having so much trouble 🤣

it says standard SDXL VAE

This was one of the first good realistic models (while retaining the Pony concepts fully), there are better ones now but many used this model for their MERGES.

This checkpoint is TRAINED!

what are you talking about. everclear is one of the best.

@potowax203 which ones are better in your opinion?

sounds like a skill issue, I train loras and generate images on this site, have not spent a penny and I'mm constantly hovering around 500-1000 buzz WITHOUT even trying to get buzz by launching loras or something like that

@bubusauce no i mean that creation-service for users dont have the local hardware ...

What are some of your common settings? For me, I love this checkpoint (v3) as is with no embeddings and no loras. But I use Euler (not Euler a), 20 steps, CFG scale somewhere between 10 and 14, (when stretched bodies happen) use a 1:1 aspect ratio often at 1024x1024, and no image quality prompting except for the standard Pony prompts like "score_9, score_8_up, score_7_up" and a negative prompt starting with "score_1, score_2, score_3, score_4, score_5, score_6, tanlines". (That tanlines is a personal preference thing for me. I don't find them visually appealing.)

Although lately, I've been having some good results with DPM++ 2M SDE Karras, 20 steps, CFG scale of 5.

Again, no embeddings or loras. And I get great results.

The only lora that I sometimes use, if I go for closeups on skin is the Goosebump_skin lora at a strenth of 0.6.

doesn't seem to be here anymore but for future visibility, fp32 models are 'full' (best for training or high end PC), fp16 is better for gens and/or lower tier PC (visual difference: imperceptible)