Using "imminent kiss" sometimes helps if there is more than one character. Prevents weird tongue issue too.

try ' looking at viewer ' (without the quotes) in your negative prompt, been working for me with only occasional looks at viewer.

@Cyberdelia Flux isn the future, first flux based uncensored model is still beta, and ultraslow(way more, than base flux).

Fire up an 5090 for Pony7:)

@Cyberdelia Yeah, totally. Flux chews through GPU like crazy. I'm on a 3080 Ti and even FP8 generation is pushing it. Thinking about upgrading to a 5080 or 5090, but man, prices are ridiculous right now

so is v88 better than later versions in your opinion ?

Really? I fear when they are supposedly worse than others. I usually just keep the new one. These damn checkpoints are too damn big and take too damn long to load up. So keep and testing them all is unfeasible.

@Cyberdelia How dare you! Nah, it's just weird that I get notifications for everything else of yours, but didn't for v10. I'm thinking maybe it's cause I left a single unread notification for v9 (to remind me to try out the other Alt version of v9). I'll give you my home address and I expect a nice postcard next time!

v10 isn't too bad, but seems to have worse adherence, worse realism and worse image quality imo. It does work pretty good with some LoRA that used to give bad results, and not as well with others. So far, my favorites are still v7.0A, v9.0 Alt-1 and some of the v6.x stuff.

Eyes are very easy to fix with loras or adetailer.

Just to add, after using it more. 10 has a tenancy to at least at times seem to be less sharp than 8.5, this seems to be what makes images look kinda cartoony at times. Colours can blend into each other more, skin can be less detailed, and for instance if I noted a woman should have eye makeup, sometimes it looks like it was applied using a brush meant to paint walls, rather than apply makeup.

Hopefully 10.1 will increase the likelihood of images appearing sharper. For now, if using version 10, it feels like using a detail tweeker, and/or increasing the weight of prompts like "photo realistic:, detailed face, detailed body, detailed skin," will need to be done to improve image results.

To users, be aware that you may need to try describing things differently than you have been in 8.5, to get the same results.

V9 is way more realistic. For semireal there are ton better models. V10 is pure semireal.

@jasongriffis394 
I would not call this a semi-realistic model. I'm not sure why, but when rendering images on Civitai, they do tend to come out more blurry, or cartoony, more often, but when I render them on my own computer, they tend to maintain a more realistic look.

As I noted in a post above, you can push it to maintain a more realistic result, using that I tend to maintain results that appear as sharp as prior versions when using it on Civitai. Then I also found, especially when rendering on Civitai, the results of 10 can be more influenced by the LORA you use with it, than prior versions were. Meaning if you used LORA trained with anime style images, you can more likely end up with more cartoony results than you used to, so would need to compensate by adjusting the weight of the LORA, and strengthening the weight of the fore mentioned prompts.

Aside from that, one needs to keep in mind that this is based on a model that is over time, with all the new models come out, getting to be viewed as archaic. If you want hyper real results, you'd likely want to use one of the newer, and more demanding models out there.

@Cyberdelia 
Meaning my joke about 11 coming out next week wasn't actually that far off.

The realistic characters/people LoRAs I train on Pony V6 XL work fine with realism checkpoints like CyberRealistic.

@SuperKinos  would u mind sharing how to proceed ? :) I never trained real people and the shots i did earlier didnt came out really good

@Cyberdelia Thanks , I will give it a shot!

Yeah it peaked at v6.6ish I think and was downhill from there in terms of both realism and prompt adherence.

what models of this series did you test, and which 2 would you suggest as your favorites?

@pink0909 only the 2 I mentioned from this series. I liked 9alt1 a lot when only that one had been released. When I tried v10 it felt like it followed the prompt more. I was experimenting a lot with realistic looking fantasy creatures, and if I try to describe the results I guess one could say that a v9 character kind of looks like if you blended the same character generated by v10 with a generic human model. Going to try v11 now.