yes thank you so much @Cyberdelia saar.

@Cyberdelia Thanks! So I guess I move forward to this v7 version, too. Time to leave classic behind 😅

@Cyberdelia Why you cant just say that it is merge of other model + some loras and tweaks? You didn't train nothing new, all data exactly the same.

@littlefluffyball What do you want to say with that? If he finetunes his dataset, that doesnt rate as merge. Be happy, 1.5 users get some new food regularly.

SD models are currently pushing up against the hard boundary of what the SD architecture can achieve. You can only pack so much information in 2 GB after all.

However, I don't think SD's potential is anywhere near tapped out when it comes to making models as useful as possible, since it all comes down to tuning for taste. And there's no right and wrong when it comes to taste. As long as you're tuning for the highest possible quality you can squeeze out of SD, there's still lots of potential to make a "better" model (according to some particular preference).

For instance, my own model Contra Maiden is in the same ballpark as Cyberrealistic when it comes to technical quality (and I'm working very hard to make the next version even better!). But which you'd prefer to use comes entirely down to personal taste. What is the default response to certain prompts, and how hard do you have to work to achieve the kind of results you're after? That's what matters.

If you just give Cyberrealistic 7.0 the prompt "woman", you very likely get a full figure shot of a nude dark-haired woman with subtly Asian traits and a neutral face expression, on a flat grey background.

If you give the same prompt to Contra Maiden 3.0 you get more varied results, sometimes full figure, but more often portrait shots, sometimes nude, but sometimes fully or partially dressed, sometimes with subtly Asian traits, but usually brunette or dark blond Caucasian, often with a slightly horny face expression, sometimes on a flat grey background, but sometimes in an indoor or outdoor environment (v4 is a lot more environment focused, but v5 will again be much closer to v3).

While you can get crisp images out of both these models, the question is just what you prefer! Do you like the more predictable output from Cyberrealistic, or would you prefer the more chaotic and unpredictable output from Contra Maiden? Do you want different seeds to produce slight variations on your prompt, or do you appreciate getting wildly different results, as long as they're all good?

Both are high quality models, and both will keep getting more polished in a technical sense for as long as the authors keep working on them. But there's not really a quality revolution to be had here, the 2 GB limit set by SD is a hard ceiling, the neural network isn't getting any bigger! Future development is mostly about tuning the response to the author's preference, and of course polishing away the last remaining issues, getting fewer logical and anatomical errors (and getting hands right more often!).

@Cyberdelia I'd be honored if you checked it out, your models have always been the gold standard for me! While I make my own models to serve my own taste better (sexy girl focus, plus a preference for highly random output with little prompting effort required!), I really admire the flexibility of your models. They're as close to "general purpose" as SD models get IMO.

I'm a bit jealous of you though TBH! You started this while SD was the latest thing, and you built a following back then, which just can't be done today. When you release an update you instantly get 20x more downloads than any newcomer, and your model instantly gets added to civitai's generator, so it keeps getting advertised and will easily end up at 100x downloads.

Meanwhile, there's not a chance in hell one of my models will ever be added to civitai's generator. When I drop an update, it gets 100 downloads, and that's it. Two weeks after release it's buried so deep in the flood of merges no-one will ever see it again.

Not saying you don't deserve the attention - you absolutely do! - it's just a bit frustrating to know that no matter how good my model gets, almost no-one will know it exists.

@Cyberdelia I wonder if it would respond to a separate t5 and clip prompt. With 32 it might. I downloaded it so I will give it a spin.

Just to chime in uninvited, two things to think about:

1. F32 models are cut down to F16 when loading them, unless you start up with a specific flag activated, which is not the default, so you won't notice a difference at all unless you do so.

2. F32 could theoretically edge out F16 in visual quality, if the model was extremely fine tuned in F32 mode. But models are very likely tuned loaded as F16. So the very slight additional resolution you get from loading them as F32 might not even be better, since that's not what they were tuned for anyway.

So my suggestion is to use F16 to generate. F32 versions are really only useful for merging, making sure you get less rounding errors in the final F16 result. I provide F32 versions of my models for that reason. It's just a courtesy to mergers, most people might as well save the hard drive space.

@contrarian This is great info I was starting to dig into how to use them better and the flag was something I did not know (using comfyui for most of my work). Also good to know I can stop looking to compare them now and just get the smaller FP16 ones.

No. I spoke with @Cyberdelia and they confirmed that V5.0-inpainting is the last inpainting model that will be uploaded. Apparently the increased model complexity with each subsequent release makes it harder and harder to create a well working in painting model, as the checkpoint is getting further from the original SD1.5-inpainting model.

The previous version in painting versions work pretty well, but if you need a better more up to date in painting checkpoint, you will have to look models other than cyberrealistic.

@ribeso8011375 Do you have any recommendations?

@Cyberdelia OK, clear, thank you. Any insights on the yet unannounced CyberSuperDuperNeg?

@Cyberdelia 

Thank you sir for your reply.

I have just started learning comfyui.

I would be grateful if you could give me complete guidance because by searching the internet and YouTube for the method you mentioned (Ultimate SD Upscale together with ControlNet). Everything is clear for automatic 1111, but I did not find an acceptable result in comfiui because there are different workflows, each of which has different problems.

Without exaggeration, I must say that your model is excellent in every way compared to other models, but if you are kind enough to prepare a training article for enlarging the images produced with this model, you will have completed the model. Why? Because when the audience sees the images created with this model that you have placed on the website (which are as good as the images of the flux model), their expectations from the model will increase and the images produced by them should be of the same quality. But there is a gap and that is the excellent upscaler that of course you have.

Sir, I won't say more than this and thank you again for the excellent model you have made available to the public. Finally, if possible, please provide a tutorial on upscaling images.

Thank you in advance.

@Cyberdelia thanks! very appreciated

@Cyberdelia  I'm so excited

@Cyberdelia Any progress on Inpainting 6.0? Can't find it.