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.

See the full description on the original page