Renderotica

Even though Z-Image and Flux handle regular text just fine, it’s still worth knowing which specific keywords are available to achieve specific effects. You can simply incorporate them into the text or list them at the end as a keyword list.

Sometimes one simple keyword can make the difference.

Here’s a link to one of the many online cheat sheets:

https://sheetly.org/chea...stable-diffusion-prompts

More complex than that, particularly for erotica.

Flux -- the base model, released by Black Forest Labs -- doesn't do X rated, doesn't respond to a prompt like "blow job"

To get that, if you want to do it through prompting, you need a fine tuned model (or a LORA) which has that particular "skill" and associated keyword.

So with erotica like from the getPhat series (models for Pony, SDXL, Flux, and now I think Krea); that's got its own list of X rated keywords; they are model specific. 

That's why you'll find that prompts mentioning spanking may not do much of anything . . . if its not trained on the concept for that particular _version_ of the model, the "keyword" doesn't mean much. You can still get spanking images by using other techniques, but if the keyword isn't associated with a visual concept in that particular model, thn it doesnt do anything

As a strained analogy -- in some of the DazPoser models, there are no genitals. You have to have appropriate geometry -- either in a built in model, or in a geograft or similar -- if you want hardcore. AI tools are analogous, if you want X rated output you need somemething that's trained on X rated concepts (though there are "tricks" that you can sometimes use to get erotica out of, say, Google Gemini or ChatGPT). 
 

Edited by user Tuesday, September 1, 2026 4:29:28 PM(UTC)  | Reason: Not specified

No, topspin is absolutely correct. Even with the so-called "uncensored" models you will find it one hell of a job to generate decent erotica without the use of LORAs in the chain. Z-Image, in particular, has a dreadful understanding of human genitalia and is very good at having a penis floating in mid air or attached to a gender that you didn't ask it to be attached to. And that doesn't even touch on the whole issue of actor positioning which is a minefield. I actually gave up with Z-Image for just that reason. CyberXL and Pony do a much better job. In fact, having a locally installed AI is good if for no other reason than to learn and work through all those issues, without flushing expensive generation credits down the toilet with an online site - if you can even find one these days that isn't "Morality Locked" or requires your identity documents to be uploaded and shared with every scammer on the planet.

What really amuses me about AIs though, is that on YouTube now they are advertising for people to learn how to use AIs like Claude, but the whole point of an AI is that it does things for you that are too difficult or time consuming for you to do yourself. If you have to learn how to interact with an Artificial "Intelligence" then it is not an Intelligence at all. It's a programming language.

Welcome to the future.

And here is the proof: No LoRAs, no nothing, out of the box with a simple prompt. All genitals are in place.

https://www.renderotica....orturl/753766/Test-Image

(More descriptions: click the image)

Of course, this model has been specifically retrained for NSFW content. Just like all models labeled NSFW.

As for VRAM: 16GB is enough for most models. Even 8GB is sufficient for the smaller Z-Image or Krea models. The model I used for testing in my last post runs just fine with 8GB,

Hmmm. I'm not a fan of that Z-Image model so far. It is dang fast, I'll give it that, but I'm not sold on the image quality. So far, for me, it hallucinates a LOT. Took ten tries to get the sample image and I'm not really happy with it, desk is in a dumb location and the overall soft look are not a win. I got the Krea sample in one try. Same prompt, same seed.

To be fair, I have been focusing on only Krea 2 since it came out, so that may play a part, but to my eyes the Krea image just looks better. And there are no toes coming out of her pussy. You can't unsee that.

Sample images: https://www.renderotica....753849/Zimage-Krea2-Comp

Edit, forgot to mention that both are using a character LORA generated from the same source images. The Krea2 render follows the look of the LORA much closer than the Z-Image image does. The only deviation from the LORA in Krea is that the  hair color isn't dark enough.

Edited by user Friday, September 4, 2026 6:46:57 PM(UTC)  | Reason: Not specified

As for Krea, I'd try out this model:
https://civitai.red/mode...m?modelVersionId=3084226

My gen AI experience so far had been focused on Wan 2.2 I2V and Krea T2I. I am dipping my toes in Dreamlight3D's AIRE render, using a Qwen image edit behind his Daz add-in. Getting mixed results with that, but think it is a prompting issue. AIRE has the benefit in that you can use your Daz characters with an AI boost and they stay pretty faithful to the features and body they have in Daz.

I use that LORA you linked in pretty much every Krea workflow. I use between 2 to 5 LORAs depending on what I'm doing. I usually use a two step Ksampler workflow, with an upscale node between the first and second sampler. 8 steps in the first sampler, 4 in the second. A 2k image takes about 2 minutes on an RTX 5060ti. I can't take credit for the workflow, it comes from Pixaroma. I also use workflows from Loratech, but those are behind a paywall and probably a lot of folks won't be interested in that.

My two biggest issues with Krea is that it can sometimes be stubborn about ignoring character LORAs, especially if I'm wanting to use two at the same time. I have a succubus LORA with pointed ears and without fail all characters ears end up pointed when I use that one. I've given up fighting Krea over that and just do a quick generative fix in Photoshop. It does a good job rounding pointed ears but sucks balls at pointing rounded ears. My second issue with Krea is that it insists in adding light into the scene when I"m trying to do something dark. I wanted to do a scene with a woman on her bed, lit only by the laptop screen in front of her but Krea wasn't having anything to do with that. Kept adding in an extra light. Might fiddle with Z-image and see if I can get darker stuff to work there.

Maybe Krea3 will fix this.

The sample image I posted used a z-image LORA and Krea LORA, both trained with the same image set. Both at 1000 steps, with a 1024 image size. Same prompt on both:

Realistic cinematic photograph. Realistic facial features, hair, and skin textures.

jenni, a woman

An 18-year old woman, completely naked, lays on her back on the bed, on top of the bedding, her naked body fully in view, her knees spread wide apart, reaching between her legs with one hand and rubbing her vagina, the other hand behind her head. Only one hand behind her head.

A small college dorm room with a single bed, white sheets and cream comforter twisted and unmade, desk with study clutter including textbooks, pens, coffee mug, laptop and papers, window above desk with afternoon golden light, potted succulent on windowsill, hardwood floor, organized clutter, photorealistic, warm lighting

Have to enforce the naked concept with Krea. I may need to get my head around z-image prompting, but I was kind of turned off about how many bad results I got out if z. Extra hands, two hands behind head and a third random hand between her legs, and the best, toes emerging from inside her pussy. Not great.

I don't really understand what the point of such a connection to DAZ is.
About two years ago, I tried enhancing Poser-generated images with GenAI. It works, sure. But you don’t need a direct connection for that.
First, I tried it with OpenPose. That works well. You pose the character in DAZ/Poser and then export the pose using a tool.
Sometimes that doesn't work so well. But then you have an online editor for OpenPose (which you can also use locally). With it, you can basically move the character back and forth just like in DAZ/Poser.Shortly after that, OpenPose was able to generate the pose directly from finished images.

https://openposeai.com/

Ultimately, I don't use that kind of thing anymore. At most, I still use images as templates. I then feed those directly into GenAI.

InvokeAI is really great for creative work. In the world of GenAI, it’s about as complex as Photoshop is for painting (okay, not quite). But once you’ve gotten the hang of it, you can do the most amazing things (e.g., copy a part of one image—say, a chair—into another image). It works with layers, similar to Photoshop.  

 

Edited by user Saturday, September 5, 2026 2:33:31 PM(UTC)  | Reason: Not specified

You're very welcome.

Yes, it's hard to keep up with the rapid developments in GenAI.