Do those hyper loras work in hi-res fix too?

@arnorwing Yes, works great.

Thanks for the tip, after tweaking the settings here and there a bit, it works like a charm without any drawbacks in quality but ways faster, great for testing prompts 😃👍

I'm still testing all three, but think I like v6.3 better than v6.4 and v6.5 so far (and v6.4 better than v6.5).

Image quality seems to improve slightly with each of those versions, but prompt/LoRA adherence and expected results gets a little worse with each revision. Still all great models, though.

@Cyberdelia Thank you, sir!

How do you merge Magic and Cyberrealisitc? I'm new to stable diffusion and have also been using both Magic and Cyber. I prefer magic for most things but cyber seems to do better photorealistic eyes. does merging get the best of both worlds?

@AI_captain Theres a checkpoint merger in SD, but honestly go to chat gpt and ask "how do you merge checkpoints in stable diffusion" and it gives a really detailed tutorial lol.

https://huggingface.co/cyberdelia/CyberRealisticPony/tree/main

Yeah that's usually what images look like when people don't know what they are doing and just use the onsite generator at 25 steps. 99% of the pictures you see posted on here are garbage. 6.4 is the best version of this checkpoint with 6.5 being a downgrade imo.

@AlleyCa7 Could you explain more, why 6.4 is the best for you. I tried all of them, and 6.5 has the clearest eyes out of stock. Without adetailer, hires. Practically i dont need adetailer, i only use hires rarely.

I have exact opposite experience. Its fuckin good out of stock, i dont need adetailer, and barely use hires.

I have ultraclear eyes, that is, what matters, isnt it?

Those old quality prompts still work. (8k, ultra highres), i use these prompts too: p1g30n, 1g0r, s1_dram. These nonsense words give ultrasharp, clear result. And you can use: amateur, raw, photorealistic, real life, photo

Edit: use 2m Karras, 6cfg, 30steps, for best realism.

Then you were lucky. I have a generation that has completely different eyes between 6.3 and 6.5. Granted it is a "cowboy" shot, so zoomed out, but with stable seed it should reproduce a similar result, so I have to assume that the changes were too big since the background was completely and the pose was slightly changed.

@Raininosi Yes, for me 6.5 has the best eyes. I use oldschool quality tags, proper negative, and the eyes are crispy clear. I dont need adetailer, and i am thinking to drop hires too. A simple upscaler enough(i use Topaz Gigapixel. In Topaz simple upscaling is 3 second, hires for me about 1.5 minutes).

example prompts? I have no issues creating hc

@Cyberdelia Thanks for the answer, It’s really helpful to know that it’s not unusual to keep older models around. I think I'll try both to see which one gives me the best results in different cases. Having more options to choose from depending on my needs sounds great. Thanks again!

Never. Always save the old models. You never know when it might get taken down because of censorship.

Also sometimes older models will produce something different than the current model.

Its the most balaced realistic model. Its full blood Pony, everything works, what works in PonyV6.

You wont find any better, because they lack in various territory.

What is your prompt for 2 distinct characters? They must be equally detailed. Asking for good sex scenes.

please provide more details. what other realistic pony models have you tried?

@isr431 I already mentioned 3, did you not read my comment fully? If you understood that it beat those two, then the other models should be irrelevant. But, every models I've tried include: AnterosXXXL, BigASP, bigLust, cyberrealistic(Pony and SDXL variants), EpicRealism, GoddessofRealism(Pony), JuggernautXL, Lustify, PonyRealism(Pony), PornMasterPro(Pony), Realism byStableYogi(Pony and SDXL variants), Realism Engine, RealVisXL, SDXXXL, Uber Realistic Porn Merge (Pony/SDXL hybrid). These models include every model that's specifically oriented for photorealism and/or NSFW. Of them, bigASP/bigLust/Lustify are cited on Reddit as the absolute best photorealistic models, but to me, the prompt adherence of Pony realistic models beats out everything. As such, all Pony models supercede non-Pony, that's why I didn't list them each. Of all the ones the Pony NSFW that were marketed for photorealism, it goes CyberRealistic > Yogi > PornMasterPro > PonyRealism >> other pony models >> SDXL/Flux models. (In terms of both exact prompt adherence for 2 people overall and photorealism for Pony models. Realism itself may be beaten, but not enough for it to make using non-Pony models worth it, unless you inpaint. But I was referring to 100% raw txt2img one model.) It was hard picking a winner, but cyber wins in terms of consistency and reliable quality potental from all my testing,

@parleyelder342 They aren't equally detailed in many cases. You may have to overemphasize the "harder" character to get, whether it's dispreffered by the model in terms of age (old or young in particular), gender (male), feature/position/action, etc. A vanilla, adult, female, is so generic that you can get away with no parentheses. While harder to generate characters often need upwards of 4-5 sets of (), lots of negative prompts, multiple synonyms, etc. My skeleton is basically like: person 1 having sex with person 2 in X position, descriptor1, descriptor2, descriptor3... where descriptor is a one word or few word phrase about either character or the setup/position. Mainly using as generic words as possible in danbooru tags style. Also don't forget to lead with score_9, score_8_up, score_7_up and negative score_3, score_4, score_5, score_6, etc. Missionary and doggy are pretty easy, as is handjob/blowjob. What are you looking for specifically? DM or reply me of what you want and I can give you examples.

@ygmdir Am not OP but advice would be appreciated,

Main features I'm after are male x female, both bodies in full view (preferably both faces too), not POV, not looking at viewer, visible penetration, bonus points for anal pen that doesn't look iffy (bonus bonus points if any idea how to actually make this less 'realistic' looking while still appealing, ie if this is a 1 in realism and the OPs semi-real Checkpoint is a 0.5 I'm after like a 0.75 if possible).

I've hardly touched realistic-style models yet (thanks for the breakdown confirming this one does indeed make the others pretty much irrelevant) but vs just Pony V6 with a couple style LORAs etc. am struggling a lot more with making decent sex scenes featuring the above.

@onelongwinter0451859 It's not that hard folks... Example of a positive prompt: score_9, score_8_up, score_7_up, naked man and woman having sex, missionary, penetration, side view, whole body, photorealistic, beautiful eyes, beautiful smile. Negative prompt: score_1, score_2, score_3, score_4. The prompts "side view" and "whole body" and optional negative prompts like "out of frame" emphasize a side view missionary. You can also experiment by combining tags like "side view missionary." Alternatively, you can go full danbooru and do something like "score_9, score_8_up, score_7_up, male, female, missionary, sex, naked, side view, whole body, photorealistic, penetration, beautiful eyes, beautiful smile." Using this particular method I get images with very high quality and reliability. I usually do 20 steps normal + 15 hi-res steps @ 0.55-0.65. Get creative... It's not hard to get exactly what you want. See here for examples https://civitai.com/posts/9827803?returnUrl=%2Fmodels%2F443821%2Fcyberrealistic-pony

same here, I think using the scores is still better on generating realistic image.

Why dont you choose a semireal model?

you want either a xinsir controlnet guided upscale with an upscaler of your choice or, if you can run it, flux, which is hard to beat but needs a lora as soon as naughty bits are involved. it is a long winded topic tho. noise injection is really helpful as well.

@jmueller thank you for the info, i will look into it.

that's a lot of complaining for things you get for free without any effort

I am utterly amazed every day by what this technology can do, and it is not even 5 years old yet. Just wait, and be patient, because I've seen so much improvement since I started using generative AI, and I cannot wait to see what it will do next.

Eh. Just do img2img 0.4 strength with a good XL model when you are done and there you go.