<h2 id="model-overview">Model Overview</h2><ul><li>Model Name: Krea 2</li><li>Version: v1.0</li><li>Release Date: June 22, 2026</li><li>Model Type: Text-to-image diffusion model</li><li>Architecture: Diffusion Transformer with 12 billion parameters</li><li>License: Krea 2 Community License</li><li>Release Format: Open-weight release and Krea-hosted product integrations</li><li>Model Developer: <a target="_blank" rel="ugc" href="http://Krea.ai">Krea.ai</a>, Inc.</li></ul><h2 id="quantization-matrix">Quantization Matrix</h2>same seed/prompt <a target="_blank" rel="ugc" href="https://civarchive.com/posts/29420066">comparison</a><ul><li>FP16 (Half Precision)<ul><li>Element Size: 16-bit (2 bytes)</li><li>Storage Size: 100% (Baseline)</li><li>Accuracy Retention: 100%</li><li>Target Cards: All GPUs (Native)</li><li>RTX 3000/4000 Series: Runs natively out of the box with zero conversion overhead.</li></ul></li><li>FP8 (Standard)<ul><li>Element Size: 8-bit (1 byte)</li><li>Storage Size: ~50%</li><li>Accuracy Retention: ~99.5% to 99.9%</li><li>Target Cards: Ada Lovelace (RTX 4000)</li><li>RTX 3000/4000 Series: Requires software upcasting on 3000 series, causing minor speed drops.</li></ul></li><li>INT8_convrot (INT8 Convolutional Rotation)<ul><li>Element Size: 8-bit (1 byte) </li><li>Storage Size: ~50% </li><li>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. </li><li>Target Cards: Any NVIDIA GPU with INT8 Tensor Cores (RTX 3000 series and newer). </li><li>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.</li></ul></li><li>MXFP8 (OCP Microscaling)<ul><li>Element Size: 8-bit + 32-block scale</li><li>Storage Size: ~50% + scale overhead</li><li>Accuracy Retention: ~99.8% to 100%</li><li>Target Cards: Blackwell (RTX 5000)</li><li>RTX 3000/4000 Series: Runs via software scaling layers; expect slower speeds due to legacy hardware limitations.</li></ul></li><li>NVFP4 (NVIDIA 4-Bit)<ul><li>Element Size: 4-bit + 16-block scale</li><li>Storage Size: ~25% to 28%</li><li>Accuracy Retention: ~99.0% (Within 1% of baseline)</li><li>Target Cards: Blackwell (RTX 5000)</li><li>RTX 3000/4000 Series: Runs via software emulation (ModelOpt/TRT-LLM); generations are slower without native Blackwell block-math pipelines.</li></ul></li></ul><h2 id="model-family-and-release-checkpoints">Model Family and Release Checkpoints</h2>This model card covers the Krea 2 model family, including the following release checkpoints:<ul><li>Krea 2 Raw: Base release checkpoint, prior to additional post-training and fine-tuning.</li><li>Krea 2 Turbo: Post-trained release checkpoint with additional fine-tuning and distillation.</li></ul><h2 id="capabilities-and-intended-use">Capabilities and Intended use</h2>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.<h2 id="out-of-scope-uses">Out-of-Scope Uses</h2>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.<h2 id="training-data">Training Data</h2>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.<h2 id="safety-measures">Safety Measures</h2>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 <a target="_blank" rel="ugc" href="mailto:safety@krea.ai">safety@krea.ai</a>. 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.<h2 id="risks-and-limitations">Risks and Limitations</h2>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.<h2 id="license-and-outputs">License and Outputs</h2>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 <a target="_blank" rel="ugc" href="https://www.krea.ai/krea-2-licensing">License</a> for more information.<h2></h2><h3 id="links">Links</h3><ul><li>Product: <a target="_blank" rel="ugc" href="https://www.krea.ai/image/k2">krea.ai/image/k2</a></li><li>API partner: <a target="_blank" rel="ugc" href="https://fal.ai/krea-2">fal.ai/krea-2</a></li><li>Enterprise: <a target="_blank" rel="ugc" href="https://www.krea.ai/enterprise">krea.ai/enterprise</a></li><li>Moodboards: <a target="_blank" rel="ugc" href="https://www.krea.ai/moodboards">krea.ai/moodboards</a></li></ul>Krea 2 is licensed under the Krea 2 Community License Agreement. For more information, visit <a target="_blank" rel="ugc" href="https://krea.ai/krea-2-licensing">https://krea.ai/krea-2-licensing</a>.

Krea 2 Turbo Official Comfy-Org Checkpoints (Krea2)

krea2_turbo_bf16 - This is the official Comfy Krea 2 Turbo BF16 checkpoint. Accuracy Retention: 100% Source: <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/Krea-2">https://huggingface.co/Comfy-Org/Krea-2</a>

krea2_turbo_bf16.safetensors

krea2_turbo_bf16

krea2_raw_int8_convrot

krea2_turbo_int8_convrot

krea2_turbo_fp8

krea2_turbo_mxfp8

krea2_turbo_nvfp4

Krea 2 Turbo is the open-weights, distilled checkpoint of Krea 2 - Krea AI's first foundation image model, trained from scratch with a focus on how an image feels rather than just what it contains. Turbo is an 8-step build distilled from Krea 2 RAW and optimized for fast, high-quality inference at high resolution.
Originally released by <a href="https://www.krea.ai">Krea AI</a> on Hugging Face. All credit for the model goes to the Krea team and the contributors listed below. Civitai is hosting a mirror so creators can generate with it on-site - please head to the <a href="https://huggingface.co/krea/krea-2-turbo">original repo</a> for weights, the technical report, and to support the project directly.
<h3>Built by</h3>
Krea AI, with Sangwu Lee, Erwann Millon, Le Zhuo, Matthew Newton, Andrei Filatov, Abhinay Devarinti, Dazhi Zhong, Avram Djordjevic, Gabriel Menezes, Will Beddow, Titus Ebbecke, Mihai Petrescu, Owen Fahey, Gian Saß, Felix Gil, and Victor Perez.
<h3>8-step distilled for speed</h3>
Turbo bakes guidance into the weights, so it runs without classifier-free guidance - Krea's recommended settings are 8 steps at CFG 0.0 with a constant timestep shift (mu) of 1.15. It generates at 1k to 2k resolution and supports outputs up to 2048x2048, making it the build to reach for when you want Krea 2 aesthetics at a fraction of the compute.
<h3>RAW-compatible LoRAs</h3>
Turbo is compatible with LoRAs trained on Krea 2 RAW, so the standard workflow is to train on RAW and run inference here. You can stack community LoRAs on top during on-site generation.
<h3>On-site generation</h3>
You can generate with Krea 2 Turbo directly on Civitai and add LoRAs on top. Weights are available upstream on Hugging Face under Krea's community license with permissive use; commercial licensing is available through opensource@krea.ai.
<h3>Links</h3>
<ul><li>Weights: <a href="https://huggingface.co/krea/krea-2-turbo">huggingface.co/krea/krea-2-turbo</a></li><li>Technical report: <a href="https://www.krea.ai/blog/krea-2-technical-report">krea.ai/blog/krea-2-technical-report</a></li><li>Product: <a href="https://www.krea.ai">krea.ai</a></li></ul>

Files with SHA256: 78bbf8f4165eda19cea3cb06c78089221932a39e2eed8af9da741f942c47ffb3