I haven't figured out how to get anything useable out of v7. There's anatomy problems in every image I output.

@vongregor817 I found that using AuraFlow=3 yields a perfect result; I think the author should highlighted them.

@gzh4170135 auraflow=3? where is this? i am not sure what you are setting

@BlueB I think the node corresponds to the AuraFlow Sampler or AuraFlow Sample. I'm not sure if that's how it's translated into English. Sorry, English is not my native language.

@gzh4170135 suppose he meant ModelSamplingAuraFlow node in comfy (see default workflow provided by ZIT-creators). For ZIT it's typically disabled by default

I have the same problem. Changing the shift doesn't help anything for me.

I would not be bored, I'm trying to learn. I feel like every single dataset we ever run for NSFW has to create the penis and vagina and penetration and barefoot tensors basically from scratch.

Is it just quantity? I've got about 150 captioned and spun, but the vaginas look like bad grilled cheese. I've got another 150 downloaded. Am I to get captioning? Or pruning to the best 80?

@chesthole I guess you are talking about training? sadly I don't have enough knowledge about it yet , I am still just a User :) .. but I can tell that you can for now go for the best 80 , bigger quantity may have effect on the time & power to process, & most cases a good suitable dataset is important, but a good set of 80 could do the job. & I wish other experts here can say more & correct me if am wrong.

I am using SwarmUI. If I copy the DL link and put it in the download model utility, after interrogating the CivitAI API, it turns to the v6 ID. Same if I input the link you referred. I see the DL link contains the correct ID. Just the interaction CivitAI API <-> SwarmUI turns it to the v6 ID. Dunno which side the problem is.

🤣 LMFAO. I heard this comment more than I read it. "She do only body horror". Haha, I needed that one.

@alcaitiff I was actually going to ask if you could zip up and share your prompt folder? You have some great prompts in there!

My character lora's are made on ZIB and then I use them on ZIT. Set my character lora on str 1.5 and the mystic lora to 0.6. Works perfectly.

I highly recommend you use this great custom node for ComfyUI asset allows you to merge LORAs together:

https://github.com/ethanfel/ComfyUI-LoRA-Optimizer/tree/main

Also maybe check out this one as well for more control of the LORAs

https://github.com/shootthesound/comfyUI-Realtime-Lora

I hope this helps.

@merchaseltd321 Do ZIB LORAs work well with ZIT like on flux? Just curious

@AnimaXx I've tested both options, they don't work well ((Or I'm dumb. Do you have a workflow that would perform this function fully? So that character lora + mystic would eventually produce a good result?

@AnimaXx All my character lora's were made on ZIB and work with the Mystic XXX loras. No loss in character likeness.

@merchaseltd321 This is quite interesting actually because I've used a ZIB Lora with a ZIT Lora together just normal stacking and it works pretty well however when I use two ZIT LORAs to it doesn't work well unless I use a custom node like Lora Optimizer that merges both of them together. Ones that allow you to change the layers of the Lora.

@merchaseltd321 holy shit this actually improved the results a f ton

Agreed. ZIT feels like a dead-end compared to ZIB.

Second this

I would like to know, what exactly you changed or did, to get better results. Because I have a simmlilar feeling. Even with the base ZIT model, this LoRa seems to ruin more than it improves my pictures.

@daMatt As you can guess from my comment my friend I am new to this. I can tell you the name and link of the Model, the text encoder and the VAE I used. Besides that I do not think I changed anything. Because I do not know much technicalities, I just follow a few guides. Please go to this link, you will finds links to downloads. You will also see why I used them. https://www.reddit.com/r/StableDiffusion/comments/1sxhiyc/comment/oipa044/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button

@daMatt And for the last two days I have started using ComfyUI, which is som much better than Forge Neo. All LoRAs work way much better in ComfyUI.

@alcaitiff Bullshit, I tried that exact workflow.
And by the way, it's not at all apparent to everyone that the images contain the workflow, might wanna do a kindness to people and specify that in the description, not everyone is an AI wizard like yourself. But of course you have no interest in doing that.
The images were generated with z-image-turbo only, which is pretty good for XXX by itself.
You're scamming people.

@alcaitiff I wish to add that it works, and very well too, thanks for your work!

@alcaitiff this is better

@lucidobscura so was your problem user error then?

@recoil7625 I am moderately savvy with these things, I ran Flux, Chroma, ZIT, SDXL and the plethora of anime derivatives, I used loras, I used controlnet with canny and depth with various models, inpainting, outpainting, face-swapping with PulId, you get the picture, but I couldn't get this lora to work. Could it be me who's doing something wrong? It's not impossible but personally I doubt it.

@lucidobscura After trying this myself it works perfectly. So yes. It's user error on your part.

@lucidobscura I've been using this LORA for quite a long time, initially on the author's workflow, later I slightly redesigned workflow for myself, recently I've revised my opinion about version 7 and use it only. Everything is working very well. There was a time when I had problems connecting LORA, because I don't know much about comfyui and the author of lora helped me a lot. No need to write that the author is a fraud, this is nonsense.

@BoyNextDora Why, if it isn't BoyNextDora, the person who literally posts all the images "made with this lora" on this site. None of which images even remotely "XXX", and can very well be made with ZIT alone. Whose exact workflow I tried and it couldn't render a dick to save its life.

@lucidobscura If you had tried his workflow you'd see that it uses regular bf16 Zimage Turbo, regular text encoder and regular VAE. Nothing out of the ordinary. I can't get this lora to work in my own workflow either, and I'm not going to download two custom node packs to try the author's workflow, but calling it a scam or fake is out of line.

I'm kind of new, too, and I've left some comments on civitai that I have since deleted. It can be frustrating because the people who make these checkpoints and loras and stuff don't always explain things for beginners. But we shouldn't be abusive toward them.

Anyway, if a model or lora is insufficiently explained and/or requires a lora with subgraphs and custom nodes that's not worth the headache of trying out, rate it appropriately and move on.

@potatometer350 Of course, he could have just asked for help, but he chose the path of negativity..

I've been messing around with this for about an hour. The best combination of sampler/scheduler I've come up with is res2s/bong tangent (what the author used in his workflow) and dpmpp sde/beta. Just about any other combination I've tried so far leads to way more mutilations and just crappy images.

I ran 20 generations with res2s/bong_tangent. 10 of them had mangled or weird looking penises or some other deformity that were insta-deletes. Of the other ten, six times the women had no nipples or weird nipples. Of the remaining four, I didn't like three for other reasons (tattoos, strange hair, weird pose). One image in 20 was a keeper.

I think it's a zImage issue. It just doesn't want to do this, and I don't have time to fight with it when there are better options. I still use it occasionally, but I do so knowing I will have to make a bunch of generations and then take the few that I like into an editing workflow or photoshop to fix.

you might be doing s/t wrong. You could start with comfyUI's ZIT txt2img template, then plug this lora in before the model sampling aura node

This lora, doesn't work for most users, because the basic Comfy UI setup doesn't have res_2s sampler and bong-tangent scheduler. Considering the amount of complaints the author should at least create a detailed guide how to make it work.

not sure what you mean, I'm getting great results with my character loras.

@Niko3DX I paired it up with +20 different LoRAs (from different people), tried a few different workflows, they all produced unusable images. I only got acceptable results when I removed any other LoRA and used only Mystic XXX. Many people agree that there's something weird with this one, I just don't have the expertise to tell what it is.