Join The Tinkerer on Whop to get early releases, private pages, and the Tinkerer Discord role - all in one place. 👉 Join on Whop
CyberRealistic is a flexible, easy-to-use photorealistic model built from nonstop testing, custom blends, and just the right amount of chaos. The backstory? It’s honestly a wild mashup of different checkpoints, but what really counts is what you get in the end: sharp, expressive, clean renders that just work.
It’s tuned for both textual inversion and LoRA, so it’s great for anyone from total beginners to hardcore prompt wizards. If you’re making portraits, messing around with new styles, or just want a model that steps aside and does its job, CyberRealistic won’t let you down.
⚙️ Suggested Settings
Sampling method: DPM++ SDE Karras / DPM++ 2M Karras
VAE: is already Baked In
Sampling steps: 30 Steps
Resolution: 512x768
CFG: 5
Upscale: 2x
Upscaler: 4x_NickelbackFS_72000_G
Denoising strength: 0.3🧠 Prompting
Negative prompt examples
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, blurry💾 Backup & Resources
backup location: huggingface
☕ Support the Project
If this model helped you hit your vision faster, cleaner, or just plain better — consider [buying me a coffee]. It keeps the updates coming and the experiments wild.
💡 Need Better Prompts?
This custom ChatGPT was made to top-tier prompts just for this model:
🔗 [Try it now on ChatGPT]
⚠️ Friendly Warning
This model can generate mature content. Use responsibly. Respect laws, platforms, and people.
Description
Model for inpainting and outpainting.
Make sure that the model name has the ending -inpainting
FAQ
Comments (24)
Will you be uploading a fp32 version as well?
Been using this model as a base when training loras and I think fp32 helps produce better models.
Unfortunately, V4 is giving me worse results than V3.3... I had to downgrade. :(
Agreed - using the exact same settings and just switching from 33 to 4 it went from crystal clear to a messy blob...
@shadow0 @Yammiekin It's a pity that you perceive it this way. The last two versions were adjusted to generate more realistic results easily. But it seems you have different experiences. I appreciate your feedback and will work on it.
And you don't have to downgrade, you can also see it as an alternative version. There are users who are still using version 1.3 or 2.0. These versions provide a different kind of results that users appreciate.
@Cyberdelia I understand that as a creator who spends a lot of time and effort on the model that is the way you see it, but seriosuly, some of the basic DNA of the model was lost in this transition and maybe you should have a second look into what happened between V33 and V4.
This model used to be really good at rendering realistic humans in somewhat unusual situations, like cosplaying as a generic example. Its capacity to do so was significantly nerfed. Overall clarity of the model was decreased as well, at least as far as human subjects are considered.
@shadow0 I hear you and really appreciate your feedback! Will compare several version with 'unusual situation' and try to see what's wrong.
Did you also tested CyberRealistic Classic? What is your experience of the latest version?
Interesting. I think I’ve actually gotten better results especially in some scenarios. The details and textures I’ve found to be a step up. Interestingly though it does seem to struggle with basic anatomy fairly consistently.
My testing shows 3.2 is the best, followed by 3.1, then 3.3.
@Cyberdelia Hi, just to add - V33 was and is hands down the best model I have ever used. The results from it continue to astound me, especially combined with my own trained LoRAs, which look so lifelike its unreal. In contrast, V4.2 has strange oddities the others do not. I get white blobs on the edge of faces and arms, eyes look crushed or flat, and the images produced no longer resemble the models I have trained on. I can run the identical seed in both and V33 gives me a beautiful photorealistic image, and V4.2 gives me an AI generated image.
I hope this is helpful feedback and is not seen as being ungrateful, as honestly V33 is just mindblowingly good, and in my opinion should be the standard all others aspire to, and you should be incredibly proud of what you achieved.
@Yammiekin you could maybe try to retrain one of your V3.3 based loras on V4 and look if you may get even better results, or if the lora compatibility in general got janked by V4. cant really accurately judge a ckpt based on loras trained on other checkpoints, even if they just differ by version number.
btw, much love to all of you for keeping on improving SD1.5 based checkpoints<3
@JarJar737 Most of my LoRas were trained on 3.1 and they work till 3.3. However, I have seen 3.2 produce the best looking results overall. The same LoRas produce off-colored images with v4, but I'm really happy with 3.2 and don't plan on migrating to 4 anytime soon.
Currently, I am working on CyberRealistic Basic. It goes back to the basics (similar to 3.2). What you see nowadays (and 4.0 is an example of this) is that checkpoints are filled with a lot of Loras. I understand the choice for that, but with CyberRealistic Basic, I want to move away from that. If you want a specific style, you can always use a Lora, but then you have the choice. I expect that this version will be ready by the end of the week or the beginning of next week. Probably as a completely new model, but possibly just as the successor to 4.0.
@JarJar737 My LoRAs were trained on the base 1.5 and just run against V33 or V4 - for some reason V33 just produces amazing results :)
@Cyberdelia Thanks for the hard work.
@Cyberdelia Any recommendations for the setting? what sampling method for example?
Hi, nice to meet you again, but please upload the full model
Another great checkpoint!
Really loving this model so far. In my experience it has been a modest step up in quality. I’ve found it’s especially improved conceptually meaning more accurate results. I’ve also found it to be fractionally faster than other models for some reason. Not really sure why, but I’m not complaining about it either. Its biggest shortcoming however is it seems to struggle with basic anatomy at times even with strong negative prompts. Often there will be a third arm or leg sprouting out of nowhere. It’s similar with clothes sometimes. For a dress, skirt robe, cape, anything flowy it also often attaches additional pieces. A robe for example for like a wizard may be resting naturally, but then there is also a piece of it that looks like it’s billowing in the wind.
I've added a comparison sheet for you to help demonstrate the issues I am getting with V4+
Hope this helps (and is meant as constructive feedback only, not a critisism)
Just want to update to say all the issues I had have completely gone in the 4.1 update, and its even better than the v3 I was using now. Thank you so much!
This is my favourite model right now, but I also agree with others that V33 is better than V4, I would really recommend going back to it and working back from. Thanks for the work though, it's definitely my favourite and I promote it whenever asked!
I have a new version (based on 3.2) which is almost ready for release.
@Cyberdelia Eeek! Exciting! Dude I bought you a few coffees as thanks, I would urge everyone to do the same because the sheer quality of your work is mindblowing
In version 3.3 it doesn't seem to need vae.
I think it has the MSE vae built in.