ALL PREVIEW IMAGES HAVE WORKFLOW EMBEDDED
Some people loved V6 as it was, others didn’t — but don’t worry, you can decide for yourself which one you like better.
I’ve just finished V7! It was retrained from scratch and focuses on reducing skin noise for a much cleaner result. Even though it uses the same dataset as V6, the outputs are noticeably different.
V7 is here! I’ve also added side-by-side comparisons in the description so you can easily compare and choose what works best for you.
Unlocks XXX on Z Image Turbo
Description
Still need work but is a nice improvement from v1
FAQ
Comments (53)
This really is the best Lora atm, just makes everything better lol
Great job. The downside is that the groin area is often blurred, as if the camera is out of focus.
seems like it will change the person to European？
If so it's feature not a bug ;]
@flo11ok874 I prefer Asians.
the best Z image lora!
shut up and take my buzz
Love you man. Thanks for the comment.
It's OK, but completely ruins the face when used I2I.
I2I isnt working very well with the turbo model, hopefully when they release the base mode it should improve
Using a low strength helps iron out details.
Love it! I had been unable to use ZIT for hardcore images since the turbo model did not want to let characters interact directly. This LoRA fixes that and gives much better anatomical details. The only caveat is that as I dial up the strength, the images get more pornographic but the bodies get more typical of porn. I’m just saying if you want to create amateur milf porn with this, you are going to have a trade off.
wow , please upload it to tensort.art , i want to use it on comfyflow for my workflow , but there is already a reprinted version of it which is not good maybe thats v1 , i tried to upload your model for private use but was not working , it shows duplicate model detected. this is here : z̷ - ᴍʏsᴛɪᴄ xxx₂ | Z-Image-Turbo - LORA | Tensor.Art And this is my uploaded lora of your v2 : Mystic XXX [ZIT] - v2.0 | Z-Image-Turbo - LORA | Tensor.Art
This LORA works great for me standalone however, when I use it with character LORAs it breaks and creates weird and broken images.
same . Unfortunate almos all loras for z image can be used only ONE. We need wait maybe for BASE model not turbo
@CyberAImania Multiple sampling, using only one Lora each time
Will this Lora works with character Lora? I need best practices for ai girlfriend bot with nsfw. What the best settings: z-image turbo + caracter Lora + something for better nipples etc?
I have experimented a little bit and it seems to be working. As with most style Loras dealing with characters, go with a low strength and work your way up. I don't know enough to give you specifics, only that this Lora seems to be worth the time for experimentation.
想问下您推荐的lora强度是多少，如果和光影或者增强型的lora一起使用呢？
Strength 1 is good, you need to experiment with other loras to see what is good. I know my sliders work well with it.
I can't get this to do male anatomy...
Download one of my images with it and test using my wf, you must get the exact same result. If you got this and changing only the prompt it don't work I will help you. But to many people said it doesn't work and don't even connected the lora to the model.
Test with the v3 version
it;s decent but int snot qwen level, i wish zit listen to prompt more closely like qwen image did. is it the clip with qwen2.5 7b that make it superior?
Great work man keep it up! I really dig it's effect on composition, but it does come at the cost of loss of details it seems
I got it, download v3 and test please.
user which workflow, it did not work in offial workflow,change the face of the girl
I think we got the cause, comfyui after 0.3.75 is different and is messing the original results I'll work on a version for the newer comfyui versions.
I got it, download v3 and test please.
I found the reason, just change the sample to euler_cfg_pp!
Thx very much!
For t2i it works really well with some slight hiccups, but for i2i I couldn't make it work properly at all, something about the process fucks the lora up. dpmpp_2s_a with flowmatch barely changes the reference image, in fact, it reduces the quality as if washing the details. Euler simple adds great detail zit is good for, but then the anatomy gets completely fucked. Penis for example turn into arms and such. I couldn't figure out how to i2i properly sadly.
what are you trying to do for i2i? anime to real? If so, start with the denoise at .30 and move up .10 for testing. Also, if doing anime to real, no lora needed.
@Melodic_Possible_582589 Yeah that. But to turn to photo-real requires big denoising, like .6+ and then the structure starts changing too much. Also, without this lora genitals and any sexual context gets lost with zit
@valkenweiss hi, I have only experienced the structure change once I went to .8 and above. Are you still keeping the prompt and just eliminating any words that say "anime, illustration, etc.?" I have done this on the side as I noticed that some non-realistic checkpoints adhere to prompts better. So I use the checkpoints and do an i2i in z-image for the realism. Remember that z-image likes long descriptions. If the anime image doesn't have a prompt you can use the limited daily free use from fluxproweb website to get some prompts. You can run it a few times to see which prompt you like.
@Melodic_Possible_582589 That's mostly how I'm doing now. Qwen edit anime-to-real lora. (qwen is goated, it makes the picture perfect structurally, hairstyle, clothes etc, but a little washed and smooth skin), then a pass with zit. Problem is that the v2 lora wasn't working properly at all in i2i. Either didn't change the image at all at low denoising, or changed too much otherwise. The scheduler and sampler affect if the lora works at all too, and I could only get the standard realism with euler simple
@valkenweiss the lora is not required in i2i. I'm using webui classic with z-image-turbo checkpoint. load your image to i2i, make sure to use the z-image checkpoint and adjust the denoising.
Hey man love your work. But i cant use this lora at all the outputs are completely mangled. I ran your exact workflow from the the models image gallery but still the result is mangled.
https://civitai.com/posts/25342811
MAN this one gave me a lot of trouble to debug. And I'm not on the bottom of this issue yet.
First I was running my own WF to get the exact same result and after I executed yours. My own WF did not returned the same result! WTF. I tryed several things until I started to suspect comfyui was the problem, I started do regress versions and the images using Z started to change! WTF²! And after a lot more tests I got the version that gives the exact result I posted in my examples. ComfyUI version: 0.3.75
I noted that this version does not load some lora layers and this lora works a lot better on this version. So I'll try to mute this layers and test on newer versions. Also, even with a lot of work I can't reproduce your result, nothing I do get me this result.
If you want to get my exact results now you will need the 0.3.75 version, if you can wait or don't want to downgrade comfyui, wait and I will come with a solution on v3.
I got it, download v3 and test please.
@alcaitiff thanks a ton man for your efforts....!! your work is a blessing to the os community!
Very cool lora, I only use it)
And again - great job!
Bro has some type of secrect forbidden knowledge because why your loras looks so much better than all the others?
Every time I say how I do it people say I'm stupid and this is the wrong way.
@alcaitiff It seems like V4 is only a matter of time.
@wyxzddsjj919 DM to me every prompt that don't work and I'll improve the dataset and retrain.
can't wait until we get the full base model this turbo model is so unstable lmao
Turbo model is fine. You can always use the de-turbo model ostris has if you really wanted.
Details
Files
Mystic-XXX-ZIT-v2.safetensors
Mirrors
Mystic-XXX-ZIT-v2.safetensors
Mystic-ZIT-v2.safetensors
mysticxxx-zit-v2.safetensors
Mystic-XXX-ZIT-v2.safetensors
Mystic-XXX-ZIT-v2.safetensors
Mystic-XXX-ZIT-v2.safetensors
Mystic-XXX-ZIT-v2.safetensors
zt-mystic-XXX-v2.safetensors
Mystic-XXX-ZIT-v2.safetensors
Mystic-XXX-ZIT-v2.safetensors
Mystic-XXX-ZIT-v2.safetensors