This is the Z Image Turbo port of the https://civarchive.com/models/2060471/rope-bound-flux and https://civarchive.com/models/2145507/rope-bound-qwen
Helps create a person bound with colored ropes, in several positions. The ropes can be around the torso, around the wrists and over the legs of the subject. It can also help create a gag (although a gag LoRA is mostly recommended for that, at least with a low strength).
A recommended format for the prompt is :
[A person] bound and gagged [position]. [Description of physical appearance].
Her arms are bound behind her back with [color] rope. Her wrists are bound. Her torso is bound, [above , below and in between] her breasts. Her legs are bound together at the ankles, the knees and the thighs. She is gagged with [type of gag].
[Clothing description]. [Scene elements and settings]
Most images in the training set contain:
Ropes around the breasts (above and below breasts - a few in between breasts),
Ropes binding the ankles and thighs or just above the knees (the word legs also works).
View points are supported (side, front, above, and combination).
Several gags are within the set, but not enough for the AI to learn gags in detail. But, colored tapes work fine.
You can then try cleave gags and cloth gags. For ball gags my samples were less than acceptable. But I would propose, that you need a separate gag LoRA for everything but the tape gag.
Now... My own personal review (v1).
Z Image Turbo is new for me, so I can't be sure I trained it with correct parameters. I used an online trainer on another site (because I tried several times on civitai and failed) using 100 images (the limit).
Sometimes, it works OK, but the Qwen version produces way better images (even with 5 steps and the lightning LoRA). The Flux version is also better. But all these come with a cost, while ZIT is quite "cheap".
Try and do send comments.
Update v2
Still needs work. Can produce better results, but still not there...
Description
FAQ
Comments (7)
How do you feel about the training and results of training on concepts like this on Z-Image compared to the other models you've done similar loras for?
In a short sentence: VFM.
Qwen is still my choice for images.
Flux Krea is my choice for realistic images and emotion (or I have to work more on Qwen).
But Z Image Turbo is pretty decent and the images are quite good given the resources needed, although the knots and the ropes are not as accurately depicted (at least in this training). But still, this is my very first training.
Heh, I came here to ask you this and ninja'd by jellai.
Other prolific lora trainers are complaining the site's training is buggy the past few days (iirc one example in nanashianon's recent article), they theorize the ZIT loras may be inferior as a result. So there's hope ZIT may end up (a bit) better.
I'm hoping ZIT doesn't completely replace ILL as the community choice, I've sunk too much effort into ILL's model and prompt faves. And for me the realism of ZIT over ILL isn't worth it, when i have Flux/Qwen for that (despite VFM, yeah).
@ravemry9 I really think when the Z-Image base model comes, SDXL will be as done as SD 1.5 (which is, not entirely, has a TON of loras/checkpoints, but will end up being used by a niche of users). But Z-Image Turbo is really only being trained with hacks right now due to being distilled, and you can't really make a good checkpoint with that, or mix loras. Still, it's almost scary how much of an avalanche of loras are coming in, as fast as we can even refresh, even with those limitations. That tells me the base model is going to be very quickly cemented as a standard.
@Jellai Interesting, thx for sharing that, 'cause I've been wondering.
I guess it makes sense, you're right the loras and community focus is coming in hard.
2 things, tho: Is it true that base model ZI won't be open source? 'because if so, that may end up being a serious obstacle; so much of SD's dominance was thru the community-wide model training that I don't think will be practical on online-only training.
Also: ZIT doesn't seem to recognize character tags. I hope I'm wrong about this. But if it doesn't, maybe it's missing more too -- like, danbooru/pony NSFW concepts or even the danbooru SFW shorthand that so many people rely on. Which may stop people from jumping on-board? I dunno.
@ravemry9 The team that made Z-Image reached out to the NoobAI team, and formally asked them for their dataset. I think even if NoobAI doesn't give them the dataset, it shows that the Z-Image team knows the importance of those characters, and art styles. They are going to do something toward that. Maybe in the full base model release, maybe in Edit, maybe in version 2, or a branch model like Pony was for SDXL, but they have shown that they want to seriously do a large training pass for characters/art styles, and maybe even hentai specifically, considering the dataset and nudity available in Z-Image Turbo, but it would likely be tame if so.
@Jellai That's cool info, thx. (you're getting me excited...)