A versatile photorealistic checkpoint
Built for high-quality artistic and creative image generation.
It delivers strong results across a wide range of styles and subjects, with excellent anatomy, lighting, skin detail, and overall coherence. Fully capable of both refined SFW artistic work and high-quality NSFW content.
Recomended Models (civitai does not let me order them)
Why SDXL and Krea 2?
These are the most flexible models that can handle both SFW and NSFW without losing too much detail. I tried Flux, ZIT, and similar models — nothing came close to the results I got with SDXL and Krea 2. So I maxed out my SDXL model and began working on Krea 2.
NOTE: INT8 version is a bit faster on older GPU's the output is recognizably different, but people and scene in general looks similar to fp8 (the output is better or worse than fp8 based on the seed and prompt).
Recommended Settings
SDXL
Steps: 40
CFG Scale: 5.5
Sampler: exp_heun_2_x0_sde
Scheduler: gits
Resolution: 832 × 1216 (2:3)
Krea 2 Turbo (Recommended over Full because it is faster and looks better IMO)
Steps: 8
CFG Scale: 1.0 (after a bit of experimentation this value can be set between 0.925 - 1.075 and result can be better or worse)
Sampler: exp_heun_2_x0_sde > euler > euler_cfg_pp > euler_ancestral_cfg_pp and LCM shows promise
Scheduler: sgm_uniform > beta_1_1 > normal/simple, and maybe ays_30
Resolution: 832 × 1216 (2:3)
Krea 2 Full (Alternative option)
Steps: 20+ (very slow so higher means more time to generate)
CFG Scale: 4+ (not sure here, same for Sampler + Scheduler)
Resolution: 832 × 1216 (2:3)
Krea 2 Advanced
Text Encoder (Abliterated): https://civarchive.com/models/2731465/qwen3-vl-4b-abliterated-comfyui-krea-2-text-encoder-bf16-fp8?modelVersionId=3070870
VAE (Clear): https://civarchive.com/models/2151326/qwen-imageclearvae
Note: You can use the text encoder and VAE either separately or together, depending on the result you want. The text encoder changes the output significantly (use with caution), while the VAE mainly affects contrast which can be beneficial or undesirable depending on the image (I would probably avoid the VAE for photorealistic amateur-style images that are intentionally meant to look low-quality and low-contrast).
Krea Gone Wild (1.0-2.0)
Play with this (if you have it) to get a different output when all else fails to produce what you want.
Comparison FP8 (left) vs FP16 (right) model versions
License
Description
FAQ
Comments (20)
One of the best Krea2 models, which I've gone through. Good job, brother.
I was reluctant to invest a lot of time in a new model but when I've seen the first results even with minimal fine-tuning I was like I won't sleep until it's done ✅ Krea 2 just became my favorite top ;) model after SDXL. Required quite an amount of programming (custom scripts) to tune it but it was totally worth it. This is not an Average Joe's merge, I even managed to keep original models metadata to give credit to the fp8 default based model I used. Next I will think how to abliterate the text encoder and a better VAE
OP, what settings are you using in your sample images? I would really help if you can describe the scheduler and sampler in the images above for Krea 2 (max-nsfw)
Default in SwarmUI, Euler + Normal
Fanstatic model! keep up the good work mate!
Thank you!
Nice model, congrats. Would you consider adding an INT8 version since comfyui can use it for faster generation.
Looking into it
Made the conversion uploading now ;)
23:40:21.998 [Info] User local requested 1 image with model 'Krea/PrimeLustXL_v1_Krea2_NSFWmax_mine_int8_convrot_mixed.safetensors'...
23:40:46.826 [Info] Generated an image in 1.02 sec (prep) and 23.08 sec (gen)
23:41:05.156 [Info] User local requested 1 image with model 'Krea/PrimeLustXL_v1_Krea2_NSFWmax_mine.safetensors'...
23:41:35.098 [Info] Generated an image in 1.01 sec (prep) and 28.75 sec (gen)
A few seconds faster on GPUs with fp8 acceleration and probably way faster on older cards, output is different as expected, sometimes better than fp8 (sometimes worse depending on the prompt).
Uploaded enjoy!
Can you please upload the fp16/bf16 version. Creating custom convrot int8 with baked in loras from fp8 is a really bad idea...
I only have fp8 it is the base model, fp16 to me is big and slow and based on results not really worth the extra space and compute.
"creating custom convrot int8 with baked in loras from fp8 is a really bad idea..." Yes and no. The wrong order can introduce repeated rounding. The better approach is to bake the LoRAs first, then perform one final INT8 ConvRot conversion.
Is it possible to download fp16/bf16 of your wonderful model somewhere? Unfortunately I can't use quantized versions on my setup yet.
I wonder why U cannot use fp8 SafeTensor its not the problematic GGUF format exactly for performance and compatibility reasons and my intuition says you could run this without issues.
@CrHasher My setup is a Mac. On Macs, we use DrawThings for generation. Other tools are too slow and unoptimized. Currently, none of the quantized versions of your model work in DrawThings. If I get a full, unquantized version of BF16, I can quantize it myself and potentially make it work in the DrawThings App.
Uploading the fp16 version! It was quite an adventure to make. First, I ran out of VRAM, and then running out of physical RAM crashed my OS. I had to put safeguards in place to allow processes to be killed before a crash and also activated kernel-level restarts. Compressed memory couldn't keep up with the barely compressible tensors, and swap wasn't working as expected, so I had to reconfigure my entire OS to handle the VRAM -> RAM -> zRAM -> disk swap pipeline correctly. But the end result is actually great, so thank you for this.
Uploaded you can test it its pure fp16
Once again thanks for the INT8 update mate! (about 40s FP8 vs 30s INT8 on my RTX 5060Ti 16GB)
Eagerly waiting for your Krea2 max nsfw version update, Bro.