Join The Tinkerer on Whop. Membership gets you early releases, private tools and a bunch of extra stuff.
馃憠 Join on Whop
馃挰 Join the community for support, free tools and early news on Discord
CyberRealistic Classic is a more unrestricted version of CyberRealistic that offers the same freedom and flexibility. It serves as an experimental platform for testing new features and concepts that haven't been integrated into the original CyberRealistic. Essentially, it functions as a fully operational alternative edition of CyberRealistic.
Recommended Settings
Sampling method: DPM++ SDE Karras / DPM++ 2M Karras
VAE: is already Baked In
Sampling steps: 30 Steps
Resolution: 512x768
CFG: 5-7
Upscale: 2x
Upscaler: 4x_NickelbackFS_72000_G
Denoising strength: 0.3-0.35Example Negative prompts
watermark, text, censored, deformed, bad anatomy, disfigured, poorly drawn face, mutated, extra limb, ugly, poorly drawn hands, missing limb, floating limbs, disconnected limbs, disconnected head, malformed hands, long neck, mutated hands and fingers, bad hands, missing fingers, cropped, worst quality, low quality, mutation, poorly drawn, huge calf, fused hand, missing hand, disappearing arms, disappearing thigh, disappearing calf, disappearing legs, fused fingers, abnormal eye proportion, abnormal hands, abnormal legs, abnormal feet, abnormal fingersless tokens:
lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurryDescription
"Starter" edition of CyberRealistic. Employ it for your personal mixes or just use it.
This model is completely operational and usable, although there may be instances where you encounter some unintended outcomes.
I use several TI in the examples which are not (yet) released.
FAQ
Comments (25)
Hey .. great Idea to have a testlab.
After several tests and experiments, i have a "feeling" that there is a "bug" / clip error in the history of the cr V 2/3 strain. .. I gave it the name "DarthVaderDesease" ;-) .. If a model has this desease there are mostly the bad eyes / mouth effects are very dominant. .. The prompt to test is in the Post-Follow Up .. The effects are very often conjoined and wrong twisted bodys.
Your semi-realistic model does not have this effect. Also the HyperV1 Model - (also here in civtai dowloadable) which has a special bad-eyes clipfix (according to his model readme) did not have this effect.
May you can take a look and may i am complete wrong. (And yes - if someone thinks i dont know how to negative them - .. i know to do that :-)) ..) but its a badly effect which also occurs in other prompt combinations.
Think I have fixed this now!
Will upload a new Beta version soon
How are you getting your image results? Are you using extra Loras or a different checkpoint to the one downloadable here?
I tried copying some of the prompts from the example images, and I get totally different faces, and all of them topless!
A few of them do look similar though.
Mmm, I don't use some extra magical things like hidden Lora's etc. Did you use the pruned FP16 or 32 version. And could give me an example?
at dooglesmie : Also the factor GPU Card Type can be a mayor factor for a different Result. Since the May Release of the SD-Webui, this can also be selected. Its called the "Random number generator source".
This was before this release, always the GPU, and this differs often between the Vendors. Also the CUDA and Transformer Versions can be a factor. Since now, this "Option" can may help to get comparable / exchangable results. Automatic11111 wrote to this new Option:
"Random number generator source. Changes seeds drastically. Use CPU to produce the same picture across different vidocard vendors."
Hope that civtai will also display this information (which is also now saved in the creation data, if the cpu was the source) in the Image Informations.
@joge25聽thats not how that works fam
@axicec聽?? what did you mean .. i am daily exchanging many seed numbers between different gpu farms (vast, google cloud instances, runpod, etc.), since i am using the CPU as the random seed genaration source, i always get the same result. This differs massive, if i use the GPU as RNG. ...
@Cyberdelia聽It was copying the prompt data for this image:
https://civitai.com/images/870069?modelVersionId=77605&prioritizedUserIds=6357&period=AllTime&sort=Most%20Reactions&limit=20
It does have a (milalc) entry in the positive prompt, maybe that could be what I don't have for mine, is that a textual inversion, or possibly the "looking_at_viewer" prompt?
This is the image I get when using the prompt from the above link:
https://imgur.com/a/3SRR7Bp
I'm running on Auto1111 with the aZovyaPhotoreal_v1.safetensors[7d3bdbad51] checkpoint, downloaded from the page showing the above image, on a RTX 3080 FE 10GB card.
@DoogleSmile聽you can't expect the same picture if you're using a different checkpoint? You said you're using aZovyaPhotoreal_v1.safetensors, the image you're copying prompt data from says '[TESTBETA-3] CyberRealistic_V3.0-FP32'
@boorad Sorry, yes, It was CyberrealisticTestLab_v30TESTBETA3 I was using to test. I accidentally wrote the name of the one I had loaded currently when writing the message last night.
@DoogleSmile聽Ah. milac is almost certainly an embed for Milly Alcock, there is an embed on this site but it isn't called that. Maybe try that?
Milac is from https://ko-fi.com/bionicstudios
@Cyberdelia聽Ah thanks.
did you notice, (also) download this? <lora:epiNoiseoffset_v2Pynoise:2>,? it's in the positive prompts.
Its more about 3.0 final, than the 3.0b3 ..
Thanks again for those two great releases. I tested the last 24 hours mostly the 3.0 final and if allowed, here my first (and second) impressions.
3.0 final feels very different than the 3.0b3 and earlier releases. Eg. more different face-types, and a much more sensitive reaction on prompting.
The age-threshold seams also different. My 20+ Personss looks now more 16 or lower.
A great step is the much lesser bad mouth / bad eyes results. (big thumb smiley !!)
There seams also the "default saturation / contrast" very different than 3.0b3 or lower. And here comes my problem:
The gap between the prompted subject (skin saturation, contrast, hue, base-color, reflection, etc.) and background is very high. (the 3.0b3 was - as i wrote - oversaturated in my results, now its very undersaturated).
If i now try to denoise, contrast, prompt to get a much more balanced / realistic composition, its getting weird. Cfg-s over 9 and denoisers are producing in my 3.0final than very often "wild flesh", overpumped vagina, flying penises, etc.
And it is still not very balanced.
So .. now i have the problem .. Using the 3.0b3 with the bad mouth, bad eyes (3.0b3 seams better in the eyes than 3.0b2), or the 3.0 final - where i have the wild-flesh issue.
If there is a way to get a 3.0 final 2 :-) with a saturation nearly the 3.0b3, that would be great ..
Thanks for your time.
Cyberdelia did you tried a Merge with LOFI or is it already part ?
No, I didn't use LOFI
@Cyberdelia聽 try it nothing to lose except that it might shift attention bias more towards asian
but we already ended up in that bias now anywys
Btw people should realize that SD 1.5 adapts based on the culture a lot of things with a simple prompt it sets the context for virtually everything.
Place,Cloth,Architecture based on the cultural heritage
@openfire聽I know Lofi and I like it. But I use really only a few other models in my mix. I use mostly personal mixes etc. But I agree with your statement about culture etc.
@Cyberdelia聽
Results we will never doubt they are real
https://civitai.com/images/923113
https://civitai.com/images/677847
His/Her/They previous results
https://civitai.com/images/128221
i think sooner or later we will have trained and tuned everything on this new standard
@openfire聽Absolutely! I purposely turned off the eyes restoration function for a specific purpose, and the outcome has been remarkable. Observing the model's behavior and its treatment of visual elements without this feature has been truly impressive.
But, you're absolutely right. The progress we have witnessed so far is just the beginning. I firmly believe that we can anticipate substantially improved results in the future, surpassing our current expectations by a significant margin.
@Cyberdelia聽
Yep they are many ways of combining different things and we just scratch the surface where in research labs it already gets done.
Currently i try to reproduce Zoyas results but of course it comes out not the same entirely and the whole high res fix behaves differently.
Interesting thing with his prompt (apart from his trained inversion not existing) i get similar results i already created once have to look in the database
even with the wrong denoising stoping at the wrong time i can see a lot of crazy details that will come through
i will post that broken result, it can happen very often when the denoising correlation doesn't work ;)
I love this imagine yourself what might comes out things :)
https://civitai.com/images/933972?period=AllTime&periodMode=published&sort=Newest&view=categories&modelVersionId=82106&modelId=57319&postId=247569
setting the upscale to lanczos fixes the problem i guess something is broken in relation to sampler and high res fix in my automatic
@Cyberdelia
If you have a short minute to read my review / problem on the 3.0 final, it would be great - if possible - to get a little response. May i am wrong with my impression, than it would also help.
The v1.0 starter / classic seems to be the same version as the 3.0final (as i yesterday checked). Is that right, or was there a problem with my download?
The 3.0beta3 is no longer for download avaible. I didnt archive them, is there a alternative download possible? It was such a great release ...
As always, many thanks for your phenomenal work.
Addendum:
According that v1 starter and v30final be the same, seems to be a problem with the identification / hash / ? of the models - or sd-webui has a bug.
Both models, side-by-side inside the models directory and runned with a x-y-z script with a checkpoint change, did not change the model. So the result was always the same.
Change manualy, the same sub-seed gives different results .. Thats weird ..
https://huggingface.co/cyberdelia
And no they are different models