My Krea2 Unlocked Diffusion Model - Int8 & FP8

FP8 Version

Updated with an FP8 version, for anyone who wants it. Same thing, different quantization. You will need a Qwen3VL 4B text encoder and either a Qwen-Image VAE or a Wan2.1 VAE. I prefer Wan2.1's VAE, it's nicer, especially the FP32 version. This is a simple merge of Krea2 Turbo, this unlocker, and this unlocker, saved as an FP8 quant, just for convenience (fewer LoRA slots occupied this way). As always, I made this for use in ComfyUI, so YMMV with different interfaces.

For ease of use, I've assembled a workflow for beginners and people who do not have (or want) custom nodes out the wazoo. It is here: https://civitai.com/models/2781539/super-basic-beginners-krea2-workflow. I made it for use with this and other turbo diffusion models, and kept it all super-simple. Anyway, on with the rest of the post~!

Int8 ConvRot Version

I like Krea2, and I talk about it a lot. People often ask me about certain aspects of "prompt adherence", which generally comes down to "will the model do boobies and wingwangs and such like, or does it require something extra?"

Well, the truth is, it depends. The Krea2 model has a couple of safeguards in place, but it was trained to be able to do NSFW stuff as well as SFW stuff without any help before those safeguards were put in place.

One method of bypassing the post-production censorship evolved into what is now called the Fedor bypass.

Another thing that works also is this ComfyUI node, which does the same thing, just probably gets there differently. Then, the guy who made it (Capitan01R) made this bypass, which works as a LoRA, just like the Fedor Bypass. Using this LoRA and the Fedor bypass LoRA seems to be the simplest way of unlocking the full potential of the Krea2 model, whether the custom node is used or not. I still like the node also sometimes, because it can create cool effects.

So, for my own purposes and to reduce my own clutter (and save a couple of LoRA slots), I merged them both with an Int8 ConvRot quant of the Krea2 Turbo diffusion model, making a Capitanized and Fedorated version of it. It frees up LoRA slots, and feels like I have one complete unredacted model, the way it was intended before the guys who made it had to bend the knee and modify it so that it could be used on their online inference platform without being overly abused by people who should know better.

Now, I know I've always put out "AIO" checkpoints in the past, because I'm one of those plug-n-play sort of people, but I can't currently make ComfyUI save a full checkpoint of an Int8 quant, and that's what I'm using right now. Int8 is really freaking fast on my computer (I have only 8GB of VRAM and this helps my potato bake twice as fast), and if my Comfy won't save Int8 quants as a checkpoint, I can live with it.

Which brings me to the other reason I didn't make an all-in-one checkpoint of this unlocked version, despite having other quant formats I could have merged. Krea2 is able to use more than just the Qwen-Image VAE. I use the Wan2.1 VAE with it myself because it's compatible and does great, and other people have even made their own cool finetuned VAEs for Krea2. If I made a checkpoint, the text encoder and VAE would be written in stone, and I didn't want to do that this time, knowing what I know now about using different VAEs. The full-sized Wan VAE especially seems to make a world of difference.

During testing I found that even with a non-abliterated text encoder, Krea2 can do pretty much anything you want it to as far as anatomy is concerned, at least, once the glue-stuck dials are unstuck. It even actually knows male anatomy fairly well without any help, something that's pretty rare to find in a generative model, and it doesn't make horridly deformed bewbies or hoo-has, either. It's practically uncensored already. It was trained on real things, and then they just glued a couple of dials so nobody could say they were being utterly irresponsible. Thank them, not me, for leaving it open enough to unglue, and thank @ftrbzmogul355 and @Capitan01R for providing the glue removers.

If you want you can use an abliterated Qwen3VL, like Huihui's, and see if it gets you further, if needed. I have no idea how far you want to push this model, and I don't want to know. If you find it isn't Fedorated enough, just add the Fedor bypass as a LoRA and crank it up some more. Capitan01R's LoRA works at 1.0, so I set it like that for the merge.

For my own purposes I find that I require nothing besides this model, @Winnougan's Int8 ConvRot version of Qwen3VL 4B available here and @Kijai's Wan2.1 FP32 VAE from here.

As for why I'm bothering to share this merged diffusion model when people can just gather and collect the same resources and use/merge them on their own, well, I did the footwork for you already, but you can still go that way. This is just for anyone who likes Int8 quants and doesn't want to have to take up two or more LoRA slots just to have a truly free and open model, lol. It's here as a convenience for those who might want it, and so I have something I can send people to when they inevitably ask me about it. I don't ask anything in return, I'm just sharing something I find useful. You might like a different setup better.

Krea2 has excellent knowledge of human anatomy already. The only need for NSFW LoRAs with this merge is if you want it to do poses or interactions that it doesn't already know (or that your favorite LoRA does way better), or if you just have a particular aesthetic preference for certain body stuff. This spares it competing with other training and having to figure out what's what, plus it's less overhead to have to load, so faster inference.

And the unlocking doesn't interfere with added LoRAs' styles or anatomy, either; it's not like duelling LoRAs, the way it can be with stacking one after another just to get a naked lady with bewbies in the same picture with a guy with a wingwang, and then having to calibrate them all to keep the image from coming out fried with 9 legs and the weener in the wrong place. I also have yet to encounter any unasked-for NSFW output, which is nice, too. You have to prompt for it.

I'm testing this still to see if I can make it better, and I still occasionally get a cartoon when I ask for a photograph, or burned areas on some body parts, but then that's always happened anyway, even before adding the bypasses and merging it. Also water from nowhere sometimes, lol. I often use very simple prompts (and make typos), and frequently use only 4 steps, so a lot of that's probably my own fault. As always with any generative model, YMMV. I'm also used to just making funny pictures of rats in tutus and vultures with hats, so there's that, too. Thank you for your patience (also, thank you to @clueless_engineer for posting all of the meat-clothed pictures, I love them. They're tasteful and low-carb, and that's always a big W).

Krea2 is 12.9 billion parameters of goodness; compared to the other models out there (I won't name and shame, they did their part in getting us here), it works fairly well for pretty much anything. This is not a NSFW checkpoint merge per se, it's just a well-trained, restored and de-mutilated universal one which is capable of both SFW and NSFW, and has been healed of its brainwashing so that it can remember itself as it was before. It's good at rats, vultures, bewbies, and weenies.

This model will do what you tell it to, so please be sure to only tell it to do things that are not going to get you into any trouble or bring hurt to anyone. You know what kind of things I mean. Don't do those things, just the good things. Be free and artistic, not gross and mean.

I'm not responsible for the things you do, either way, nor is the model or the company that provided it to the world, nor are the guys who made the stuff that removes the glue. You're responsible, so please be responsible. This model is provided for use conditionally under the Krea2 Community License Agreement, which is available here.

PS: I've included a few NSFW sample images now (you can see them on .red), so people can see what this model can do without any additional LoRAs. It's not generally my style, but I hope the samples are tasteful enough, and help. They were generated using only simple prompts, and I only had to throw two of them away, lol.

PPS: If you have any issues using a sample image as a workflow, it's because I rename models when I upload them. Look inside the subgroup that has the sampler (the purple one) by clicking the upper right corner, and you can change things inside it, like which model is expected to load. Or use this simpler workflow instead.