ComfyUI and Diffusion Pipe support Z, I think there are more trainers for training Z loras

@qek I tried training one using diffusion pipe but It doesnt load in comfy I just get lora key not loaded errors

first, but at what price. ain't no one gonna remember this (if it doesn't get updated to be actually good)

not on civit so doesnt count

@QualityControl fair enough

Due to censorship

The model has to be trained specifically on things. It has no idea what a weewee is unless it has been explicitly trained with a lot of pictures of weewees. When people say "uncensored" they mean it doesn't have built-in mechanisms to prevent weewee images. But training still needs to happen, thus LORA like this one.

@qek it's people like you that spread ignorance and misinformation. the model is not censored. it is simply not trained on dicks. unlike you.

@crombobular I mean "censorship" in the vein of ignorance is still something when presumably most of the humans on the planet have genitals

Like personally it should be able to generate perfect human anatomy, but the problem is you can only do that if you make it brain dead. SO they just make it ignorant instead. I think Flux has made people thorny around the topic, myself included, because it feels like they're stealing. They are stealing.

@llhappyll822 how exactly are they stealing by giving you a free model? jesus dude. and no. they are not the same. flux2 is definitely censored but z image isn't. this is an important distinction.

@crombobular AI in the literal sense are mostly trained on the plethora of human-generated content on the internet. It can be your opinion that these things are being given to us for free, but it's not exactly true, because AI steals from us in far more quantifiable ways than simply talking about the license they stick on the model. And that's all I'm really getting at. Any heavily-censored, brain-damaged model that doesn't come with real free licenses is probably just theft. And I was referring to Flux, really, and the knee-jerk reaction to call these models "censored".

@crombobular That's like saying a Ken doll isn't censored. Of course it is. There are different kinds of censorship and different degrees of it. Blocking images from being generated is one kind. Not including basic anatomy for nudity is another.

@ss9999 you're wrong :D

@crombobular The only reason to delete basic body parts that make males male is censorship. It's not even a debate. It's a simple fact.

@Spat1984 i tried to learn a lora with this fork too but got then these error when loading it, any clue ?

: lora key not loaded: lora_unet_layers_5_attention_to_out_0.lora_up.weight

lora key not loaded: lora_unet_layers_5_attention_to_q.alpha

lora key not loaded: lora_unet_layers_5_attention_to_q.lora_down.weight

lora key not loaded: lora_unet_layers_5_attention_to_q.lora_up.weight

lora key not loaded: lora_unet_layers_5_attention_to_v.alpha

lora key not loaded: lora_unet_layers_5_attention_to_v.lora_down.weight

For info, Ostris Ai-Toolkit seems to support z-image.
https://github.com/ostris/ai-toolkit/commits/main/

OP does mention that it was very poorly trained and also take into consideration that the Z-Image model we have not is the full version (Good training on the turbo model might bot be possible). I still have hope. :-)

Z image is mostly NOT trained on NSFW data. It is uncensored/no built-in censorship like Flux has, so once the base model comes out, we can safely train it ourselves, without refusals. I don't know how long it would take to train it on that tho. - that's as much as I know.

All of the models that you mentioned were finetunes from the models that couldn't even do half of what Z-Image can do with all its flaws, while being a distilled model officially released and not by community. Especially Chroma, it's just a large finetune of Flux Schnell, which is an extremely censored model, that it is ridiculous to even bring it up.

So yeah, you were criticized for your idiocy and lack of understanding of what community means when they say "not censored model", which doesn't translate to being a porn capable model.

lil bro doesn't understand the difference between censored and not trained, lmaooo

This is a flat misunderstanding of how these models work.

@Thijs From my testing it almost seems like a 50-50. Like sometimes it perfectly understands the erotic thing I'm trying to do, and other times it has absolutely zero idea. In both cases genitals and stuff are completely whack. However, I think because we even get that "50", it implies it's probably just very ignorant instead of braindead.

Also the point about using the qwen clip is a good one, we already know how low of a bar qwen sets for nsfw.

@munchkin Chroma is still more capable than z-img, but takes three-four times as long for an image.

@llhappyll822 More flexible, yes. Point was that the current Z-Image-Turbo is a distilled model that wasn't finetuned on a more general and NSFW data properly, unlike Chroma, which skews it towards photos and the likes and only has a surface knowledge about anatomy. Which is better than where Flux Schnell was at the time. In other words, false equivalence of different types of models.

It's literally brand new. It takes longer than 48 hours to get quality loras out. Possibly a couple of months AFTER we get access to the full model.

@munchkin I mean of course, but if your requirements need to be satisfied now, and not in 2 years when someone fine-tunes a z-img base, then okay. I guess you'll be saying then that the z-img fine-tunes are irrelevant and we shouldn't talk about them in the context of ... well anything, I guess. Idk man, the guy simply said "the models that right now, today, can make a penis". Z-Img is not that model. Yet, sure, yet. But today.

@llhappyll822 That's just BS, exaggeration, or a denial of reality. Especially when you say that under a LoRA of even a Turbo model that already has a nudity and all the other LoRAs that already come out with that, let alone the base model that wouldn't be distilled. Not to mention that rumored Z-Image anime finetune by the devs themselves. So what 2 years?

Chroma took a lot of time because the base model sucked for it and was very censored - that's the whole point. That guy over there doesn't get it and calls it censored, which can't be further from the truth. And Chroma isn't even that better, it just has some basic understanding.

What's the point of trying to make up some false words for that guy?

does illustrious really give male gens out of the box [not a finetune]?

It's a shame they didnt cook some Male anatomy but to be honest nobody has yet so let's stay positive. I want to try and train some male penis once the base model is released.

Chroma truly is wonderful for male anatomy.

@thomas852002887 Honesty is positive. People trying to claim that removing basic body parts from training isn't censorship don't know what censorship is. There are different kinds of it and one of those kinds is the deletion of body parts that every single male has when born.

@quoifroixaugullau4491619 I have frequent issues with bad hands with Chroma and other anatomy problems.

@Spat1984 Sounds great. I am looking more into the Amateur style part and would like to mostly exclude professional images. Maybe you are right and i should mix in some professional stuff.

Some Images from Reddit are really good but its really hard to get full compositions (if there are good compositions, the quality is mostly not great). I think my next steps would be looking into regularization data (which i will generate myself in the Model). Maybe it fixes the issues with my prototype.

Currently, my prototype Loras are good at the overall shape but do not get the balls or tip right most of the time. I hope to get that fixed with a few more/better images in the Training Data. My Images get changed a lot with my prototypes so i hope regularization can fix that.