V3 (bf16 & INT8 ConvRot) :
Improved skin texture for greater realism
Optimized lighting effects, more subtle and with better contrast
Enhanced painterly visual style
Corrected color shift toward warm tones, color rendering is more neutral and realistic
KREA 2 TURBO Vs KREAMANIA
V2 (bf16 & fp8) : Improvement of version 1 by merging an unpublished personal LoRAs to optimize lighting, textures, and graphic effects.
V1 (fp8) : The removal of certain restrictions compared to the model turbo fp8 scaled, along with improved realism and nudity rendering. The goal wasn't to focus on an NSFW model (even though it is capable of that), but rather to unlock the full range of creative possibilities.
NOTE : Sorry, I can't seem to get functional INT8 working, even though I've tried all the recommended methods.
Recommended ComfyUI extensions:
https://github.com/ethanfel/ComfyUI-Krea2TextEncoder
Description
Steps : 8
Sampler : heunpp2
Scheduler : Normal
FAQ
Comments (53)
Thank you :) Here are loras: https://huggingface.co/collections/krea/krea-2-loras
will this work with Forge Neo?
Sorry, I can't be of any help, I don't use Forge Neo.
@Adel_AI No worries, thank you for the quick response I appreciate it
I'm waiting for them to add support in Forge Neo
Krea 2 support for Forge Neo is still on the roadmap so not implemented yet but planned to.
Roadmap · Issue #1254 · Haoming02/sd-webui-forge-classic
I have been using it with Forge neo https://github.com/Stable-yogi/sd-forge-krea2 its supported already
I have a feeling you're going to have a blast making non-stop LoRas for this model. I haven't been this excited since Flux 1 and then Klein (for editing).
You're reading my mind, unfortunately, it won't be happening right away, I'm a bit short on time.
Is it just me or does Krea2 Loras not working on these finetuned models?
Do you have an example of a LoRA that didn't work with this model?
@Adel_AI Seems it does work. i just takes about 50x longer than the normal model.
15 min per generation with this model with lora. 13 second with the normal Krea2 Turbo bf16 with lora. 12 seconds with this model and no loras.
@nisse91 You guessed it: it’s a model that takes a coffee break during inference—it really takes its time.
@nisse91 --enable-dynamic-vram
@limon7 Thank you. I found that solution too. It works now. <3
That's really great. I'm sending you my thanks with the green buzz because I don't have yellow buzzes, is bf16 difficult to make? I'm greedy because I think the quality will go up even more
Hello my friend, thanks for your buzz, whatever their color.
I’m going to try to devote a bit more time to Krea 2, because it’s exceptional. In my opinion, it’s well ahead of ZIT and gives Ideogram V4 fierce competition.
@Adel_AI Thank you for your reply. Ideagram 4 is absolutely overwhelming. Maybe it's the power of bbox and flux2 vae, as well as the excellence of models. I think I was born in a great time. I'm lucky to have so many great models and creators.
Hi! That’s strange—I thought I’d already subscribed to you, but apparently hadn’t—though I’ve fixed that now. I’ve been following your models for a long time, and many of them are already saved in my collection; Krea is no exception.
I really hope you continue developing it—Krea truly came out of nowhere and surpassed even Z-Image, which is itself quite good, yet still falls short of Krea.
Wishing you all the best with your future projects and endeavors!
Hey. Given all the issues with the Civitai site, it’s no surprise to find things getting mixed up with links and subscriptions, it’s happened to me, too.
Yes, Krea 2 is a huge surprise; it far surpasses ZIT, and I think its performance is right up there with Ideogram V4.
I hope to find the time to really dig into it.
Thanks for your interest in my work.
you are truly top contributor to a community, thank you for your effort and this high quality model! @CivitaiOfficial , change the rules and make a normal ladder to reflect peoples' output- they actually truly deserve it
Above all, it’s a personal pleasure; if others find it interesting too, then all the better. We don't need rankings to share the enjoyment. Thank you for your appreciation and interest.
we have a contender to overtake illustrious at last.
Absolutely amazing work!
One question, if this is, as it says, a BF16 version, why is it only half the size of the original? is there maybe a real bf16 hiding somewhere? ;)
That's a mistake; I'll correct it. Thanks for the feedback.
you are welcome :) but ... is there a BF16... ? ;)
@bewunderteradler112 The Krea 2-raw version is in bf16, but it is a model primarily intended for fine-tuning and LoRA training. For inference, it is preferable to use the fp8 turbo version.
https://huggingface.co/krea/Krea-2-Raw/tree/main
This I know. I was just hoping that a BF16 version, so your fine-tune of the base version with the turbo lora applied on your data, would have less visible image artifacts after generation.
but it can also be my settings so.... nevermind :D
I am almost certain there won't be a significant difference between bf16 and fp8; it will be barely noticeable. I’ve already made this comparison with other models, and it seems to me that the same applies to Krea 2.
Admittedly, bf16 offers ideal precision, but the resulting sizes already disadvantage the vast majority of local users.
another reason people might want the bf16 is that you can make int8-ConvRot from it and keep most of the precision that other 8bits versions will not have.
Is an nvfp4 variant available?
I'm not considering it for the moment.
It seems native NSFW support is the only highlight of this model
hence its name
Best model!
Its just another perfect unet. Masterpiece !!!
Thanks for your feedback
Hi there,what type of bypass did you add? And trying to do some sex scenes but sadly I'm just getting a illustrations or stuff like that
Hey, what do you mean by "bypass"?
The model is capable of generating NSFW images without any issues; there are no restrictions, nudity, pornography, and sexual poses are all uncensored.
@Adel_AI you see,I tried this prompt: A Snapchat selfie of an Asian woman with makeup raising up her shirt to expose her breast with her hard puffy nipples. The caption says "Come tit fuck me" with a winking emoji
And the only image I get is a portrait of a black Lady or an Indian girl portrait photo, idk what's happening
https://drive.google.com/file/d/1iznUIrctXfjqkos2so_vz12nMwcrpBYE/view?usp=drive_link
It generates an image that roughly matches your prompt. Check your workflow; the issue isn't with the model—there's almost certainly an error somewhere.
@Adel_AI I will upload the result I got,maybe you could check it to see if something is wrong?
@Neon_signs The PNG Adel_AI provided contains the full workflow. Just drop it using ComfyUI to see if you can replicate it. If it works, then just use his workflow.
Good day! If you have the time, could you please release an int8 version? It significantly reduces generation time.
Hey, to perform INT8 quantization, I need to reconfigure my machine to install a version of Triton compatible with sm_120 (GPU 5090); I plan to do this in the coming days.
Workflow has a Load CLIP (#56) with krea2. Where did you get it or how did you fix the Load CLIP?
Thanks.
https://huggingface.co/Comfy-Org/Krea-2/resolve/main/text_encoders/qwen3vl_4b_fp8_scaled.safetensors
I didn't fix anything; it's the standard "Load Clip" node from the ComfyUI core.
Why not lora (extracted)?
You're referring to the LoRAs used in the V2 merge? They aren't necessarily models I'm satisfied with on their own (due to a loss of consistency, for instance), but they can prove interesting in merges by focusing on specific blocks.
@Adel_AI Why you upload full model? wht not lora (extracted from full checkpoint)?
First of all, I don't see the point of extracting LoRAs from models; the resulting quality is terrible. Secondly, why are you talking about extraction? My approach involves specific merging.
I am looking to create a custom-calibrated U-Net model so that I can use it later with my LoRAs.
When I want to create LoRAs, I do it by training on a dataset.