@Cyberdelia Can confirm in my testing that this new release really does improve on age-related rendering of faces. In past versions I always added "child" to the negative prompt to rectify that quirk and after a couple of hour-long rendering sessions on this new version I no longer need to add that to my base negative prompt anymore.

Awesome job on working that out, loving this new version so far and getting great results with it! Thanks for the great work!

V21RC, V10classicstarter man i love those lol. Glad to see you still working hard on improving things. Can't wait to get back home and mess with this.

What is better and not better in your mind, and what about Cyberealistic Classic and all the versions there?

(I think I figured it out--I had accidentally switched to DPM++ 2M SDE Karras, and it seems you need a lot more steps on that sampler to avoid glitchy images.)

I got a completely blue body on one of my renders, normal face. On another one, the subject lighting was completely contrary to the prompt and the environment setting. That prompt was kinda garbage though, and I'm thinking the blue body was a one off despite being very strange.

maybe change VAE?

@crappypoo The blue body thing seems like the model resisting nudes for some reason, it looks like a rubber suit or something. Did it on both recent versions, alternately it tries to put gloves or rubber leggings to cover the subject in some way. I also use embeddings and loras in combination, and I haven't tracked what I might have been using, but no other model does it, and I'm darn sure it's been through different lora/embedding combos. Just odd, if I figure it out I'll check back.

@Onvisi DPM++ 2M SDE Karras works great on this model. It's my go to that I use on it daily actually and I've never gotten blue faces or anything like that.

Depending on the prompt and seed I usually get the best results from running 22-25 steps @ 5 cfg. Once in a while a better result will come from 30-34 steps @ 7 cfg.

A good rule of thumb for running any sampler is to start at a base of 25 steps @ 7 cfg and run 4-8 batches. Choose which image most closely represents what you want, then run that seed through an x/y/z plot first with a range of 20-40 steps set to generate 8 images. Pick the best of those and set your steps to that number. Then go back to x/y/z and set the cfg from 5-10 with a range of +0.5. The best result of that is your optimal step count and cfg setting for that seed and prompt combo.

If you're still getting blue faces or whatnot, then you've got a wonky prompt or are using Loras or T.I. or VAE that doesn't play nice with the model. Command line arguments can also affect rendering as well. For the record all I run is Xformers and I have Clip set at 1 and always have Restore Faces checked.

@MysteryWrecked Not sure why you're getting that result when trying to render nudes on this model. I render full-on outright porn images with this model and have never had an issue with it resisting nudes or sex or anything NSFW. My guess is you're using Loras to get the result and they're messing up the model's denoising ability(which is common and why I steer clear of Loras in favor of using prompting and ControlNet to get nudes or sex to render), or you're over-prompting to get the result. This model does best with a more minimal prompt approach.

The problem with relying on Loras to get results is if they were trained on some niche or style-specific model they're not going to be very compatible with other models. Especially if said model is also niche or style-specific. In the rare case I do want to use a Lora with any model, I train one on the specific model I intend to use it on to get rid of compatibility issues and 99% of the time that'll get it working as intended.

To get nudes on this particular model you shouldn't even really have to prompt it much beyond just adding NSFW in the correct position in your prompt order and possibly giving it a bit of an emphasis boost(highest I've ever had to boost it was 1.2 if I remember correctly).

For the record all I run for command line arguments is Xformers and I have Clip at 1 with Face Restoration on with CodeFormers set to 0.5 strength. I don't use the CyberRealistic_Negative T.I. as I have my own negative prompt T.I. that I get better results with based on how I like to structure my positive prompts.

Anyways, not trying to downplay your issue, but I don't want people reading and thinking they can't generate nudes or lewd renders with this model. With the proper setup it absolutely can, and tbh, in my experience it does a better job of it than the majority of models that are geared specifically towards generating realistic nudes.

@DAV3X I agree that the problem is likely on my end, and I have been running Lora in combination with TI and controlnet to some unintended effects, but I've never seen this particular problem so far, and not with any other model on the same configuration. I know that cyber can be great at NSFW, but I've been running versions of this since Feb/Mar and have seen it be less kinda prude in the past. Seems to me, that may be why Classic came out, maybe the amazing quality gains over its development came at some compromise.

I tend to ask a lot of a model, complex situations and elements that understandably push the capabilities of AI rendering. My imagination is wild, and I try to create things that you can't just dig up in a Bing search. I've been on a steep learning curve with all forms of AI synthesis, and finding the limitations of the tech and working within them has been a process. All of this is amazing, including this SD model in particular, but AI is still far from rendering my dreams into our shared reality. All things considered, that may likely be a good thing, haha.

@Onvisi turn on "Always discard next to last sigma" in Web UI settings when you use DPM++ 2M SDE (Karras or not). This should mitigate glitching of those samplers in some models. Note that this thing works for all samplers, and you may want to place checkbox for it on quick settings bar.