My Krea2 Restored Diffusion Model (Final Form)

The final form (probably) of my favorite blend of Krea2 Turbo and the best of the unlocking bypasses. This is (to the best of my knowledge, anyway) as close to the original, not-yet-censored version of Krea2T as it gets, without adding any extra training data or forcing it to hallucinate a roomful of unasked-for details. I did it solely to "improve prompt adherence".

This is a merge of the standard Krea2 Turbo diffusion model, the Fedor bypass, and the TextFusion Refusal-Reduction LoRA. It is available in 5 different sizes/quants, including BF16, in case you want to use it to make your own merges, or even if you just like using OG-sized models and have the VRAM for it (I do not, lol).

Please note: this was made using ComfyUI, and works in ComfyUI, but other inference setups may have issues with some quants. I'm working on figuring out how to make them accessible to more than just ComfyUI users, but I can't promise anything. I'm doing all of this on my own setup at home, and it's not exactly SOTA. Use a LoadDiffusionModel node instead of a LoadCheckpoint node.

Sometimes an ordinary text encoder will make this model produce water out of nowhere or do other weird things, like replace your realistic subject with a cartoon, make them wear flesh-colored underwear, or cover their skin with a patchwork of weird stickers. If you need an unlocked text encoder, try looking here. It can help a lot. You might still get a cartoon sometimes when prompting for a photograph, or get a weird skin effect, but it happens a lot less frequently.

You'll want a decent VAE too (for image quality); I recommend the FP32 version of the Wan2.1 VAE over the standard Qwen Image VAE. Yes, it's compatible (at least in ComfyUI it is), and it makes nicer hair and fabrics and stuff.

This is my personal favorite Krea2 merge. It knows to leave eyeglasses on nudes. You can prompt it for someone to be nude, then tell it they're wearing something, and they will be nude except for that specified something. It has native knowledge of certain specialty clothing items and things that other models have always required LoRAs for. More often than not you'll get normal-looking hands and feet, even. It's like an evolved SDXL with 10x-better training, minus the brain damage. Not trying to throw shade, but I haven't seen extra legs in over 1000 generations, just saying. It's still not perfect, but then what else is? It still makes nicer bewbies than any other potato-friendly model without any help from a NSFW LoRA.

Sample images were generated using nothing but this base diffusion model (Int8 for most) and simple text prompts. Some used only 4 steps, I think. And yes, I threw some away, because I suck at this stuff. No additional LoRAs were used in making the samples, just my own pathetic prompting abilities. Also, this is just a decensored base model, it doesn't have any added training for aesthetic improvement. This is two-edged; the base model still struggles with some things (like skin details), but you can also use LoRAs without them fighting against the base model's bias, like they would with a merge that already included style LoRAs, detailers, someone else's preferred body types, etc.

Naturally, your mileage may vary with different quants on different machines. I find the Int8 runs faster on my machine than the NVFP4, even though it's larger. Int4 takes me longer to initialize than Int8, but then it gives me ~1.1 seconds per step (which is pretty fast in my world). Google which quant runs best on your computer, and maybe you can save yourself a lot of downloading and testing. And before anybody asks me, no, I can't do an MXFP8 quant. My machine just refuses to do it; believe me, I've tried over 9,000 times. >.<;

And bear in mind, the smaller the quant, the greater the difference in output quality, just as a general rule. This doesn't mean smaller quants are all crap, it just means you might need to use more steps to get better images, use a realism enhancing LoRA, and/or maybe do an upscale/detail fix. Use clear, simple, specific prompts. Where there's a will there's a way, and don't let my shoddy sample images fool you into thinking otherwise.

I've lately been exploring improvements for "photography" with Krea2, and have found that I can actually get pretty realistic outputs even with the Int4 quant. One thing that helps greatly, after much searching for realism enhancers for Krea2, is the Pornmaster Realism Slider. I get very good results with it at a strength of 5.0, and it helps a lot with that random-yet-all-the-same "AI face" thing (and probably more). To improve skin detail (or the lack thereof) Loraholic's Skin Detail Slider can help a lot. When using both sliders, I set Loraholic's to 2.0, and Pornmaster's to 2.5. Some skin detailing LoRAs can cause weird effects (flesh-colored underwear on nudes, scars and skin creases out of nowhere) if set too high, so remember to start at low strengths and work your way up to find your sweet spot. Every LoRA must be balanced.

I also like to do a detailing "upscale"pass with the 1x-ITF-SkinDiffDetail-Lite model or similar (there are some more you can check out here), to help finish off the last of the plasticky skin texture and to help bring out finer hair detail.

I find that 8 steps can be way better than 12, though 9 might actually be the sweet spot, just like with some other "turbo" models. It can poop out cartoons and Pixar-style stuff and conceptual quickies at lightning speed with only 4 steps. Also lower-res photographic-style stuff.

One other little note, I have found that image quality improves if you use an empty negative prompt, rather than replacing it with a ConditioningZeroOut node. The empty negative prompt actually makes a difference with turbo models. I have tested it myself.

Krea2 is licensed under this license, and neither the people who made the model, the people who made the unlocks, nor I are in any way responsible for how you use this model. Please prompt responsibly, because even my crappy sample images can get flagged as indistinguishable from real photos sometimes, and our AI generations can have real-world consequences, regardless of how fun it is to fool artists on Reddit, etc. Be good, and have fun! I know you can do both~! XD