V3 Is almost done!
All images have WF embedded, I recommend you to use euler/beta 12 steps on turbo. It gives better skin and textures.
Description
First version
FAQ
Comments (34)
So soon ? Woah, thanks
Seems like it affect likeness a lot.
My initial tests w/ lora are promising (regarding cock quality). Good enough to pry me away from Ideogram for a bit. Seems like this model (generally speaking) might be more prone to plastic when it comes to male anatomy vs Ideogram. But I appreciate to have yet another NSFW model. It's faster and I prefer simple natural language. Thanks for sharing the lora!
Awesome😍 👍🏻
Would you train a version of Boogu Image?
Krea-2 will leave Z-image in the dust. This LoRA is 10x better compared to anything anyone managed to make for Z.
I agree with this it also understands Json prompts plus this model is very tuneable. Plus even the image to image is close to what a editing model can do even before the editing model comes out.
I usually have a good eye for artifacts and obvious AI hotspots (eyes) but this lora is very realistic and can trick the eye
Krea 2 is the most uncensored base model of a all modern models, it already saw a lot of uncensored NFSW during training, it knows all the required anatomy.
Building a LORA on top of already existing knowledge is much simpler ;)
This is cool, but it kills known likeness in the model. Forces same-face for generations.
How about v2?
Really good, but is ideogram better? These models seem similar in realism, but seems like to my eye, ideogram v4 has the edge.
Krea 2 focus was never realism but style diversity
@brahianvalles Ah. That explains it. Not my bag, but others may prefer that.
ID4 is better quality, slower, and slightly more work to get what you want. Both are new and need work. But both are reasons to delete your ZImage folders.
@brahianvalles The same is true for ideogram.
ideogram is heavy and the json thing is annoying, it's incredible for control but just for... this? i prefer flux9b and this krea2 is becoming my fav
you can use the ID4 prompt mechanism on krea 2 fyi :)
Idk what ppl are saying krea realism was not the focus, from what ive generated its very realistic more than zimage and ideogram. Maybe its the settings, this lora though is insane.
Ideogram 4 is a heavy model, and you can generate up to 5 images in krea2 while only one image is generated in ideogram. The quality is not significantly higher, except for high resolutions. Ideogram is also more focused on layout, as it requires a clear json format with the location of each detail, which can be useful for ultra-precise requirements but may be excessive for generating various similar content.
@pirkvil Isn't Ideogram 9b, and Krea 2 12b? Why are people saying Ideogram is 'heavy'?
The model is good. But it changes people's faces, bringing them closer to the same average age and appearance. However, I understand that it may be difficult to teach this in a different way.
DO you mean the model or the lora? Cause loras can cause what you're describing. Many lora trainers just use previous assets to quickly get loras out for new models, those assets tend to be images trained on SDXL based or other very common image generation models which have same face syndrome. They tend to not train on a diverse set of facial features. The model by itself I've not had any of the issues you described.
@DaddyWolfgang I'm specifically referring to this Laura. Although she is currently the best NSFW Laura for Krea 2 in my opinion, I am not eagerly awaiting version 2.0.
Hey there. What's your opinion about Krea2 compared to Zimage and Flux2klein? Thank you 🙌
Thank you! Best NSFW lora for Krea2 so far, awesome work.
why? what does it do that the others don't? and what strength are you using?
@necroryona it has many uses, it even works well on anime
@necroryona more stable is generating genitalia, mixes better with styles. I still use it, 1.0 strength, together with conditioning and Enhancer nodes. Although, new LoRAs appear so quickly that there might be a better one already.
Thank you so much for this wonderful LoRA! It's absolutely fantastic and incredibly useful.
🤝
NSFW should be an integral and natural part of any model — true creative freedom isn't complete without it.
The Lora is very good. But what I noticed, it really likes to create top-down views even when prompted overwise.
BRO! Krea2 is unhinged AF, its the most capable NSFW model Ive ever used ever and Ive been using AI for NSFW since SD1.5 days. Krea2 has nailed every single wild NSFW pose I asked for, every single one of them!. No model has ever done that just a few days or weeks into release, not even months or years after release. The only big problem is that the image quality sucks compared to Ideogram, imo ZIT has better photo quality out of the box than Krea2. I'm struggling so hard to find the right settings for maximum realism with Krea2, while with Ideogram the quality is there right out of the box. Krea2 is plagued with ugly SDXL synthetic training data. But man does it nail the NSFW perfectly. Krea2 is hands down the best NSFW model, nothing comes close to its capability and creativity, not even a closed source model like Grok comes near it. Once again, its only major downside is its tendency to create a synthetic aesthetic.
Seriously 1.5 was incredible but look what we got now 💯💯💯😜
You’ll get synthetic looks if you only use this Lora or rebalance node. Try realism engine