Model Overview
Model Name: Krea 2
Version: v1.0
Release Date: June 22, 2026
Model Type: Text-to-image diffusion model
Architecture: Diffusion Transformer with 12 billion parameters
License: Krea 2 Community License
Release Format: Open-weight release and Krea-hosted product integrations
Model Developer: Krea.ai, Inc.
Quantization Matrix
same seed/prompt comparison
FP16 (Half Precision)
Element Size: 16-bit (2 bytes)
Storage Size: 100% (Baseline)
Accuracy Retention: 100%
Target Cards: All GPUs (Native)
RTX 3000/4000 Series: Runs natively out of the box with zero conversion overhead.
FP8 (Standard)
Element Size: 8-bit (1 byte)
Storage Size: ~50%
Accuracy Retention: ~99.5% to 99.9%
Target Cards: Ada Lovelace (RTX 4000)
RTX 3000/4000 Series: Requires software upcasting on 3000 series, causing minor speed drops.
INT8_convrot (INT8 Convolutional Rotation)
Element Size: 8-bit (1 byte)
Storage Size: ~50%
Accuracy Retention: Near-lossless (~99.8% to 100%). Ranks just below GGUF Q8 but generally outperforms standard FP8 and MXFP8 by rotating weights and activations to suppress outliers.
Target Cards: Any NVIDIA GPU with INT8 Tensor Cores (RTX 3000 series and newer).
RTX 3000/4000 Series: Runs natively with full hardware acceleration on both. Highly advantageous for RTX 3000 (Ampere) cards, which lack FP8 tensor cores but feature dedicated INT8 pipelines, entirely bypassing the FP8 software upcasting penalty for faster generation.
MXFP8 (OCP Microscaling)
Element Size: 8-bit + 32-block scale
Storage Size: ~50% + scale overhead
Accuracy Retention: ~99.8% to 100%
Target Cards: Blackwell (RTX 5000)
RTX 3000/4000 Series: Runs via software scaling layers; expect slower speeds due to legacy hardware limitations.
NVFP4 (NVIDIA 4-Bit)
Element Size: 4-bit + 16-block scale
Storage Size: ~25% to 28%
Accuracy Retention: ~99.0% (Within 1% of baseline)
Target Cards: Blackwell (RTX 5000)
RTX 3000/4000 Series: Runs via software emulation (ModelOpt/TRT-LLM); generations are slower without native Blackwell block-math pipelines.
Model Family and Release Checkpoints
This model card covers the Krea 2 model family, including the following release checkpoints:
Krea 2 Raw: Base release checkpoint, prior to additional post-training and fine-tuning.
Krea 2 Turbo: Post-trained release checkpoint with additional fine-tuning and distillation.
Capabilities and Intended use
Krea 2 is a text-to-image diffusion model that generates images from natural-language text descriptions. The model is designed to support creative, commercial, developer, and research use cases, including image generation, concepting, design exploration, visual production workflows, and integration into applications and creative tools.
Out-of-Scope Uses
This model is not intended or designed for uses that violate applicable law or regulations, infringe or misappropriate third-party rights, generate or facilitate unlawful or harmful content (including CSAM, NCII, harassment or defamation), or support fully automated decision-making that adversely affects legal rights of individuals. This summary is non-exhaustive. Use of Krea 2 is subject to the Krea 2 Community License Agreement and must comply with the Acceptable Use Policy. In the event of any conflict, the Krea Acceptable Use Policy and Krea 2 Community License control.
Training Data
This model was developed using a combination of publicly available data, data licensed from third-party providers, and synthetic data generated through proprietary methods. The training data includes images and their associated captions or text descriptions.
Prior to training, data was filtered to remove certain categories of harmful content and reduce low-quality, duplicative, or irrelevant data. Krea also used curated and synthetic training data selected to improve prompt following, visual quality, and alignment with intended use cases.
Safety Measures
We implemented safety measures across the full model development lifecycle. We applied targeted fine-tuning techniques to reduce the model's susceptibility to generating harmful content in response to both direct and adversarial prompts, and we conducted multiple rounds of internal and external safety evaluation before release.
For Krea's hosted products incorporating Krea 2, we deploy input and output classifiers using a combination of proprietary and third-party detection tools to flag or block policy-violating prompts and generated images.
Because this is also an open-weights release, Krea does not control downstream deployment of the model. Under the Krea 2 Community License, deployers are required to implement content filtering measures or equivalent review processes to prevent the generation or distribution of unlawful or policy-violating content appropriate to their use case. Deployers who fail to implement required safeguards are in breach of the license. See the license for details.
We conducted multiple rounds of internal and external safety evaluations before release, including adversarial testing designed to assess the model's resilience to attempts to elicit harmful or policy-violating outputs. Testing covered sexually explicit content, non-consensual intimate imagery, child-safety risks, and other high-risk content categories. Based on these evaluations, the release checkpoints demonstrated high resilience against violative inputs across the tested risk categories.
Krea maintains reporting channels for harmful, illegal, or policy-violating outputs at [email protected]. Reports involving potential CSAM are escalated to NCMEC as required by law. Krea reserves the right to update model weights or revoke access in response to identified misuse patterns.
Risks and Limitations
Krea 2 is a new technology and there are risks associated with its use. Testing conducted to date has not covered, nor could it cover, all possible scenarios. The model's potential outputs cannot be predicted in advance and may, in some instances, produce inaccurate, objectionable, or otherwise undesirable outputs.
This model is not intended to provide factual information. The model may fail to generate output that matches the prompt, and prompt following may be influenced by prompt style, specificity, language, and phrasing.
Before deploying any application using this model, developers should perform safety testing and tuning tailored to their specific application and must implement safeguards required by the Krea 2 Community License.
License and Outputs
Krea does not claim copyright or other intellectual property rights over content generated by users of this model. Users are solely responsible for their outputs and any subsequent use of those outputs. As with other generative tools, the nature of a user's inputs influences the outputs produced, and prompts may produce images that implicate third-party rights. Users are solely responsible for assessing and addressing those risks. See the Krea 2 Community License for more information.
Links
Product: krea.ai/image/k2
API partner: fal.ai/krea-2
Enterprise: krea.ai/enterprise
Moodboards: krea.ai/moodboards
Krea 2 is licensed under the Krea 2 Community License Agreement. For more information, visit https://krea.ai/krea-2-licensing.
Description
krea2_turbo_nvfp4 - This is the official Comfy Krea 2 Turbo NVFP4 checkpoint. Optimized for Blackwell 5000 series Nvidia cards. Uses ~25% VRAM compared to BF16. Maximum VRAM savings. Source: https://huggingface.co/Comfy-Org/Krea-2
FAQ
Comments (42)
are there any image comparison between the models?
Quality is going to be pretty similar. The data in the main post compares them. It's more about which GPU and how much VRAM you have to burn.
Here's a same seed / same prompt comparison https://civitai.red/posts/29420066
I am using a amd 7900xtx and i wonder which model i should choose. Only the fp16 one?
bf16 would likely be the choice I think for you.
для генерации nvfp4
An unrelated but how is the performance of the xtx? I've heard that nvidia cards run faster but all that vram is so appealing.
@ForestGreenAI it takes about 45 second for a 1024*1536 size picture. Anyway, 7900xtx has a 24gb vram which can use most models, slower speed is tolerable i think.
ps:bseides the bf16 one, the fp8 model also works well on my 7900xtx
@bnx005514 Thanks for letting me know. I'm on a 4060 myself and while 16GB of VRAM is good for most models, it has a very narrow bus, so it's slow and I was thinking to upgrade.
@ForestGreenAI just dont think about using amd graphic card for any ai usage lol, I choose to use a 7900xtx rightnow just because I sold my rtx4090 at a really high price(
Лучшая модель которая отправит на пенсию SDXL, для генерации используйте NVFP4 , для обучения LORA только FP16
Она отправит SDXL на пенсию только при условии, если будет такой-же быстрой, гибкой, адаптивной, и без цензуры.
Но если все эти критерии действительно будут соблюдены, и (главное) удобны в использовании, тогда АБСОЛЮТНО ДА! :)
Да, но NVFP4 только если у тебя карта RTX 5000 серии?
@aiAnaista Форматы NVFP4 и MXFP8 предназначены для карт Blackwell (серии 5000); в противном случае их быстродействие будет ниже. (google translated)
А на ForgeNeo запускается?
ForgeNeo не знаю, я перешел на COMFY, 3 года назад, благодаря возможности подключать разные ноды эта самый лучший воркфлоу
@daceheg192491 Hello. Wtf is in settings of gallery Convrot version - "Some images have been hidden based on moderation preferences set by the creator" :(
@xNzX IIRC there's 2-3 folks that literally filled the entire gallery with 1 image posts. I think I marked them ignored. No other filtering is going on.
Which version is better to use on RTX3090? I read a description, and it caused more questions, than answers :D
And also, 24GB size... It it even gonna run on RTX 3090?
If you are using comfy it's great at memory management, if I were in your shoes I'd try the bf16. (edited to remove the fp8 recommendation for speed, apparently not accelerated on 30 series, may as well try mxfp8 as a fall back if you are going to lose acceleration anyways)
BF16 or FP8 (less vram, slightly slower) should work
I use FP8 on my 3090, but with high-res and adetailer in Comfy it takes a very long time per picture (500 seconds on average).
@aiAnaista Worth trying the BF16 to compare? If you do please report back
@daceheg192491 I tried krea2_turbo_fp8 and krea2_turbo_nvfp4.
krea2_turbo_fp8 generates picture in ~65 seconds.
krea2_turbo_nvfp4 generates picture in ~35 seconds.
As you can see, the second one almost is twice as fast with my specs (RTX3090 24GB + 32GB DDR4) which pleasantly surprised me.
I used Lonecat's Simple Workflow for KREA 2 for this test.
@aiAnaista NICE! That's really good. The smaller file size must be helping quite a bit which is impressive.
In comfyui, you should be able the full models with 24GB VRAM. I can run full raw and full turbo with no issue, not sure how much my 128GB RAM plays into it though.
INT8 should be ideal for 3090
I strongly recommend using INT8, the final highlight of the current Series 3 cards
If somebody can make a LoRA to make the model be able to accept multi image inputs for editing and style transfer, i will bless you for everything
Here is this LoRa. But for best similarity (if two persons transfered), you need to transfer each one-by-one, not both simultaneously.
https://civitai.red/models/2761113/krea-2-identity-edit?modelVersionId=3112922
@xNzX Lol the fact that discovered that made it even better
Based on my initial experiences, Model is generally good. I'm running it with 16GB of VRAM and it's fast enough. I normally use Illustrious as my base model. If I compare the two:
Background creation is quite good. Backgrounds are detailed and generally accurate.
The accuracy in perspective and body proportions is not bad.
Applying different styles isn't bad. However, only general styles work.
It can recognize popular characters, but not sufficiently. Illustrious is much better at this.
NSFW filter is strong. It only allows safe content.
>I normally use Illustrious as my base model.
If I may ask, why?
@New_Name I think it's easy to use. It's a bit weak on backgrounds. However, it can recognize a lot of fictional characters and styles. It's better at character focused images. It's quite successful at creating the fictional characters I want in detail with high accuracy.
It works quickly because it's not a large model.
It has more Lora support.
This model is clearly was trained on a lot of n_sfw content, but then it was trained to specifically avoid it. The model will avoid specific themes even if prompt looks safe and only composition is questionable, so it clearly knows what to avoid - that's impossible without training. Also you can see how it's easy to unlock specific nsfw abilities with small datasets which is not possible for models that were not trained on such things. Currently it's not so difficult to repair its ability after a medium training with universal dataset.
you literary need 32kb tiny lora to unlock it, actually its not just nsfw, without unlocker lora KRA2 cannot even do face expressions
My guess is that it was meant to be uncensored out of the box but in last minute they were forced or decided to censor it, so they did it in a way it will be easy to bypass and probably leaked how to do it as bypass was made on first day.
Model seems to be trained on everything.
@bitzupa Any "unlocker" lora I have seen on this website seriously destroy prompt following - reduces knowledge to an equivalent of SDXL, but with better quality. It looks like that approach is too rough and should be refined somehow to not degrade model so much.
I'm not sure about, but here is something about 'unlocked' text encoder for Krea 2:
--https://huggingface.co/huihui-ai/Huihui-Qwen3-VL-4B-Instruct-abliterated
--https://huggingface.co/huihui-ai/Huihui-Qwen3-VL-4B-Thinking-abliterated
--https://huggingface.co/prithivMLmods/Qwen3-VL-4B-Thinking-abliterated-v1
??--https://huggingface.co/huihui-ai/Huihui-Qwen3-VL-4B-Instruct-abliterated/discussions/1 -- 404
**--https://huggingface.co/mradermacher/Huihui-Qwen3-VL-4B-Thinking-abliterated-GGUF/tree/main
**--https://huggingface.co/huihui-ai/Huihui-Qwen3-VL-4B-Thinking-abliterated/discussions/2
!!--https://huggingface.co/mradermacher/Huihui-Qwen3-VL-4B-Thinking-abliterated-GGUF/tree/main --
Huihui-Qwen3-VL-4B-Thinking-abliterated.Q5_K_M.gguf -- 2.89 Gb
@xNzX it won't help much for existing model, but it would help a lot for training process (if someone would ready to spend a lot of GPU time for adaptation to existing Krea model and then training on a big dataset with around 200k images). These models have a difference in output, so Krea should be trained to adapt for output changes. It looks like default text encoder had guardrails, so descriptions for some content was not translated to diffusion model. Diffusion model was trained on nsfw, but nsfw content was not tagged well.
We need a hero, who can turn this into the new Pony. Make NSFW great again! 😃😃😃
It is on par with or better than Pony already in my opinion, if you use one of the several NSFW work arounds.
https://www.reddit.com/r/DegenDiffusion/comments/1ukitpq/krea2_comparing_various_current_nsfw_loras/
@J1B, in terms of prompt adherence krea 2 is superior to Z-image. SDXL-era models - not even worth mentioning.
But in terms of styles, characters, NSFW concepts and poses - krea 2 is still far behind pony, illustrious, even anima.
If the goal is to make generic girls in realistic style than yes, bypassing filter and a few loras on top will easily surpass any other model