Unlocks XXX on Z-Image
Description
First version, be gentle. Or not...
FAQ
Comments (25)
Oh woah, and to think i had nothing to do tomorrow... AMAZING STUFF? can't wait to try. Thanks for your efforts :)
Idk what your dataset is but it always does so well. This replaces the zit version in a good way. ZIB loras play nice with ZIT loras. Thank you for your contribution.
Dang, looks very promising! Excited to try it.
This should work on ZIT as well, yeah?
Yes! This image was made with turbo: https://civitai.com/images/120787927
nice step but looks like it needs more data, because some of angles and details are not good. Also needs more penis heads and anus pics
Hi mate, this LoRA heavily biases the output when used with Z Image Turbo. It turns almost every Asian face/body into an overweight Caucasian person, which wasn't an issue in Mystic V5. Would love to see an improvement on this in the next version. Much appreciated!
Wow, thanks for the incredibly specific and totally-not-at-all-suspicious bug report, mate!
I’ll get right on fixing that ethnic weight-gain + surprise race-swap issue… right after you send me a dataset of 100 perfectly diverse, high-res Asian girls respectfully demonstrating every single position from the Kama Sutra + bonus DLC, in crystal-clear PNG, perfectly captioned, zero watermarks, varied faces/lighting/outfits, naturally lit glory shots.
Just DM the Google Drive link whenever you’re done downloading/uploading/curating/annotating that little project. Shouldn’t take more than… what… a weekend? Maybe two if the deepthroat lighting is tricky.
Once I have that reference material I’ll be sure to retrain the LoRA overnight and drop V2 with zero Caucasian conversion bugs. Promise.
Appreciate the help.
Agreed, but i had pretty good luck with the negative prompt "asian" (*no hate at all, just noticed that too*).
Still struggles at aligning the anatomy correctly for inverted/upside-down poses. Also tends to mess up the hands and feet if they are near the edge of the frame. The details in the anatomy are better though.
The most wanted LoRA when a new base model is released. Mystic! Thank you 🙌
Don't know what I'm doing wrong, but it's not working for me at all. Outputs are complete mess
Download one of my images, drop on comfyui, use my prompt, it must be exact the same, if not you may have some problem on your config
lora need more data its like close but not enough
I really hope that we get more Checkpoints and LORAs from the community As the quality of the base model is nowhere near as good turbo model colours seem off and the anatomy seems like flux Klein 😂 nothing to do with your LORA I'm just saying in general. Hopefully Alibaba open sources Qwen image 2.0 As I think that will be the next generation of open image generation.
Was this made using the same Dataset as v5 for ZIT? or a different dataset?
If it was using a different dataset, it would be cool to see a version using the exact same one as ZIT V5 did just trained on ZIB instead of ZIT.
Honestly I feel the anatomy with this version is a bit off especially in comparison to ZIT V5. ZIT V5 had really good detailed anatomy close up whereas this one tends to cause more misshapen genitalia and bad anatomical details with the addition of a significant skin style bias that makes it look less detailed bad in some lighting. It also tends to force pubic hair even when not prompted and also when prompted specifically not to.
Either way you are awesome and I hope you continue to work on this.
The differences you noticed come from the ZIB model itself. It has more freedom and flexibility in its outputs. Because of that, it can sometimes produce unusual or unexpected results. Distilled models, on the other hand, tend to be more constrained and consistent: they reliably produce correct results but may lack the ability to explore more complex or creative variations. ZIB is more prone to issues like detached limbs or imperfect skin rendering, but with well-crafted prompts it can achieve very accurate and high-quality results.
@alcaitiff No, actually I was using the Lora with ZIT. ZIB Loras can be used with ZIT, and thats what most people them for. I can try using it with ZIB but the quality difference and generation time between ZIB and ZIT is quite significant due to the higher step requirement.
Or if you are talking about because of the fact that you trained it on ZIB, but other ZIB Loras work fine with ZIT typically.
Did you use the prodigy optimizer for this?
Had some real trouble with training base, lost skin detail etc. Yours seems to have worked well though. Any tips?
Config is not the most important part, your dataset and captions are.
@alcaitiff I think my dataset and captions are very solid. I got worse results than with turbo and same dataset, so it could be the training platform using lower resolution for base. If it's there's nothing special you do with settings, all good.
I wish I knew what I was doing wrong. Every image I generate has something wrong with it: third legs or arms, grotesque body parts, breasts with no nipples or areolas, wildly wrong ages, two foot long penises. And the women are suddenly ugly. Not average looking. Ugly. And they often look 70 years old even if I prompt specifically for a 43 year old. Prompts that work fine without a LORA have gone completely haywire. It doesn't matter which LORA I try to use. Sometimes they will generate something close to good, but there's always some glaring deformity or something. I just don't get it.
i need help.. how to use it with a Character LoRa? need balanced strenghts :)
Nice but seems biased towards heavy makeup like eyeliner etc. - I imagine this is because of the dataset?
This Lora rules. Will we ever see a new version?