This Checkpoint may create some wild stuff, very cool stuff and horrible stuff, - but it is so different to all you know. It will be exZITing!
Go wild, go nuts, or just use three words. This model will surprise you with every generation.
I’m a Civitai user since 2022 but never released anything here. But when I was experimenting with Z-Image Turbo and SDA something wonderful happened.
A mistake, a quirk, a human error made this model. It was like a slap in the face, it was like waking up inside a dream, it was completely strange and wonderful.
This model can create unlimited completely different images from your one prompt.
This is why I am posting it here. So you can have fun with it, as I have. It reminds me of the very beginning of stable diffusion, when sometime things went wrong, but sometime something very unexpected came out of a generation.
Have fun again when pressing CTRL+Enter. Oh, and it is very uncensored and mostly works with skanky stuff. Also works pretty well with loras.
Post your crazy results here, if you want to share your pleasure.
Infos you may need:
CFG: 1
Res: All (I use mostly 1216 x 1536 or 1024 x 1536)
Sampler: What you wish. Best sharpenss with dpmpp_2s_a
Scheduler: Most of them but I use Beta
10 Steps
Clip Skip -2 (also optional)
All examples are direct creations, without 2nd pass or upsampling.
More Information about the model and some tips:
The base model has been trained on about 8000 hand selected images, ~1500 are full text trained and captioned by quen3-vl-8b. All others are trained on WD14 Tags. This makes the base turbo model accepting simple prompts, and it works far better then I thought it is possible. In the beginning everything nude was horribly wrong, but got better and better as the training went on. the training data is a wild mix of images, but most contain people. The model was trained to 1000's of poses, body shapes and body positions. you will see an unreal amounts of variety with a simple prompt like: "A woman is doing a pose for a photo.", - try it (it might go also horribly wrong some times). The injection of SDA at .6 strength helps the model to make sense from the pose, but sometimes it fails. But often it works very well to do even the most complicated body "dislocations". "A woman is posing for a photo."; might do less extreme or silly poses.
Since the data contains nudes and the model does not understand NSFW or SFW, you may need to use: "A woman is posing for a photo, wearing clothes." If you want less Asian looking women you can try "A blonde woman is posing for a photo, wearing clothes.", the hair color steers the model extremely unfortunately. What also works is: "A caucasian woman is posing for a photo, wearing clothes.", but you probably know the jazz already.
Description
FAQ
Comments (13)
Man you weren't kidding, that's impressive. And that was an accident? Wow. I really can just leave a 5 word sentence for a prompt and get a wildly different picture every time. A fair number of eldritch abominations in there but still.
Testing 1 million things that don't work, and all of a sudden one thing works ;). What i still don't know is where the extreme sharpness comes from \o/. I had something like this in mind, but it never worked. then DSA fixed my idea with the wd14 caption mix and the 1000's poses. Pure luck.
Absolutely amazing model!!!!!!! Keep the updates coming...:)))
I don't know how you have done it, but this is fantastic.
Update: Nope. Did not work at all. Not sure if it is possible to make it better. Don't expect anything at this point.
Working on the new version every day since a week. This is thing is extremely hard to tweak further. Another night of baking, maybe, just maybe I can release it tomorrow. The test-image count is over 3000, and that's just hard work. I was so sad two day's ago when certain things did no longer work at all, or it went completely crazy. I had a model where everything had a layer of frost or sweat on it, no matter what prompted. o.O
What is a bit frustrating, is there sheer amount of merges posted on this site. The chance is big, you wouldn't even see a new one, that does things different. Just 100 IL models a week with the same anime stuff, uff.
Try to merge it with my model 3.5 or 4 , it has some issues but i think it will stabilize yours a little
@BastardOG Thanks for the offer, but I like the model how it is. Did you test it? I don't want it to loose it's unique abilities, so I leave it at it is. I was feeling some kind of urge to make it better. But everything I change makes it worse, boring or collapsing. For what it is, it's it may be in fact one of the best turbo model out there, not because I'm a genius, but by sheer luck of the data mix and how these damn black boxes do their thing.
Qwen is so hard censored on certain things, that I think it's not worth it to try to overcome it's flaws (nude males, sex).
Feel free if you want to experiment with a merge, it can certainly add dynamic to other models I think.
@cushycrux i do not have problems with censorship with the original Qwen... it does exactly what i prompt it (some things are still hard to get , many nsfw positions or hardcore porn movies scenes that seem impossible to do in real life too) I'll merge your model in a new model and post it when it's done .
not getting any improvements , merge not working
I am enjoying this model, but my images all seem to bright and too contrasty. Any idea? what vae are you using?
See my images for the settings. Normal flux vae is used.
@cushycrux Thanks.. got it. Love this checkpoint. Strangely the prompts in your example images are too short. They just give me noise, but if I add a few words they are fine. I'm using neo-forge. Works great though. Love the variety of images.
This is a great model, guy. Exactly as described. Very different output from other models. Sometimes misses, but hits more often than not in very creative and pleasing ways. I don't want to tell you to improve it. Maybe it's impossible. I just came back looking for a int8 convrot to see if I could get it to run a little faster.