Pretty sure that's an issue with training loras directly on the turbo model and not the base model (which isn't out yet). Not really anything lora trainers can do that I know of. Just gotta wait for the base model to drop and get these retrained.

@whatsthisaithing I really look forward to using your base version.

Nice. Thanks for the link. Glad to see people working on this :)
Another tip is to add heaps of extra details about the "anatomy" of the prompt. If you overload it with intricate details it can give almost decent anatomy without any loras. I use a separate qwen3 4b abliterated GGUF loaded from Searge LLM Node to "enhance" the prompt ;)
Can't wait for the full model to see what it's really capable of.

It gives an error in ComfyUI. Does this Clip Text Encoder work with ZIT?

@Learning2025 update comfy gguf nodes

man, your deserve the world!!!

@0xbcc8950 thx!use in gguf clip loader?dosn't seem to work

@automatal comfyui gguf single clip loader on lumina2 type plugged into the same spots on the default workflow.

@tenstrip thx!!

@wyxzddsjj919 Its a text encoder like the normal one so in text_encoders, but its gguf so in the comfy_ui GGUF node pack theres a gguf-clip-loader other than that its the same.

anyone have a comparison if this actually helps or is it placebo

@crombobular equally as horrifying, however the GGUF does a pussy slightly better than the tensor but if you're after the sex change with a coat hanger look, stick with the tensor.

very useful. Thanks for sharing.

@tomspin The model is still under trained on anatomy. But especially with the penis loras and such I've seen a big difference in how they are drawn better and details are depicted. Also you see much more gains when using explicit Chinese language prompts with phrases like 背后注意 and I think there is where you might get the real gain.

@tenstrip I'm not knocking it at all, actually i think it's really impressive that i can run such a model on my 5070ti and have such impressive outputs in seconds. I don't really give a shit about porn generation with it, I was just answering that other guys question about comparison between the GGUF/Tensor clips.

I didn't know there had been work done to run the Llama stuff natively on PyTorch, so props to you for mentioning it, i learnt something, have another heart.

Everyone who's been using q8, I added a new link to a bf16 version someone linked me that is also a straight plug in over the normal one. It should be less detrimental to quality.

@crombobular It's not going to automatically teach the model genitalia. It's about just making sure that you have nothing latently censoring your prompt or impeding generations and making sure your Loras are actually getting full effect.

@crombobular working on one now (EDIT: I found the difference to be relatively low, maybe only 5% better, but not consistently)

@Destinyfaux Neither of them have an optimal run through heretic I think but there are some prompts and seeds where the heretic will definitely unveil more nudity or directly change the pose, but sometimes it will also do abstract things. You only need it if you're trying to go after straight-up porn I think, although I find the nipples seem to come out better sometimes. It's not worth banging back and forth between them you just pick one and run it.

@tenstrip Hi, I tried your model and it was nice, thank you. I even used koboldcpp to run it and found that it is possible to write adult pornographic literature short articles, how do you bypass the security restrictions of QWEN3? I downloaded the QWEN3 uncensored model from the HUGGING FACE website, trying to get it to write pornographic literature, but it was rejected (same in Chinese), do you have a recommended uncensored version of QWEN3? I would like to get an 8B or 14B AI model that can write erotic literature😍

@wyxzddsjj919 It was run through this https://github.com/p-e-w/heretic. It comes out "abliterated" meaning refusals were removed. It was just a try to have impact on the image model so I guess that's good you can use it for that. I'm not sure where to find other ones but look for 8B or 14B abliterated or find someone to run them through the heretic.

@wyxzddsjj919 I also tried running it using kobold and was surprised that it worked because the original file refused to load. it's funny to use it to write a prompt ) And about other qwen. Both 8b and 14b have heretic version on hugging face. Search mradermacher/Qwen3-14B-heretic-GGUF. Or are these exactly the ones you've tried?

@tenstrip THX~😋

Is there anyone out there be willing to explain some basics for the less expert? I’m only glimpse the concept here. Something about the text encoder refusing to send vocabulary to the model? It would be so great to have an expkaination with more basic description of what this actually does.

@timstertimster3 More or less yeah. LLM's have a refusal mechanism but it's usually more about making bombs and chemical weapons or committing crime. The uncensored one has no real interaction with Zimage, at least the turbo model. The full model may be delayed because they are adding censorship after seeing what's been done so far and it may come in to play then.

估计得等base版出来之后了，turbo版训练的效果不好

通病

@chocoboking 之前另一个老外作者也说没人脸，但就是还会影响了，你用的什么流？官流吗？我没加任何其他lora就直接官流挂一个你这个Lora，提示词直接特定的人名，还是改了脸了

官方貌似说turbo的底模训练的lora效果不那么理想，试了下人脸的确有点不稳定

系数调低点也许可以

@geeccyolo594那只能期待ZIB模型了

我用好多ZIT的lora'都会改变人脸，貌似这个模型就有问题，但系数到0.3以下影响不大

@ssoh188 有个新出来的一线天lora我试了，不改变人脸，但影响画风和画质

@ssoh188 2026 Cute Pussy 一线天美鲍：https://civitai.com/models/2192886/2026-cute-pussy?modelVersionId=2469119，这个不影响脸

z-image_pov_cowgirl_zimage is ok

确实很完美

不过看来，目前只有躺着的情况下是完美的

@chocoboking, But it can still do non "close up" images? Is it like v1.0975, but with the additional images?

i was wondering the same thing. seems like there all just spreading from the preview images. i just want realistic NORMAL anatomy down there. not the full insides hanging out lol

You mean the Text Encoder not the VAE right?

@Learning2025 yea sorry

@nikolatesla20145 You are what's wrong with this world.

what you use cfg and steps? still get garbage...

哈哈哈哈，笑他妈死

都是他妈很刑的图，谁敢发

bro知道你很急，但是你先别急

笑死

卧槽，我也没见你自己返啊

我是个高尚的人，你不知道嘛？操！

This isn't for IL this is for ZImage which is a higher different model architecture and higher parameter model.

He is locked in. even his typing is formatted for prompting.

TM在这还研究涉不涉黄？

这TM差点把我笑死

你真他妈的破事真多。不玩滚，装什么清高。

傻逼

哲九十沙比

你都到这儿了 你给我说涉黄？？？are you kidding me?

笑出内伤了

人机

还真是

你勾选XXX内容的时候不知道自己在干什么吗？

他其实在陈述一个事实~哈哈

你清高！

推荐一个插入状态好的lora

ASK.