Skip to content

Krea AI / image model

Krea 2 Turbo

Explore Krea 2 Turbo, its capabilities, use cases and ComfyUI workflows for image generation on HexGrid Cloud.

Krea-22 workflowsExplore workflows ↓

Deploy Krea 2 Turbo with ComfyUI

Choose a workflow to explore its setup, inputs, downloads and generation examples.

01Workflow guide

Text to Image

Create images from text prompts with Krea 2 Turbo. Explore the ComfyUI setup, workflow instructions and generation examples.

Text → Image

Explore workflow
02Workflow guide

Style Reference

Guide image generation with a visual style reference with Krea 2 Turbo. Explore the ComfyUI setup, workflow instructions and generation examples.

Text + Image → Image

Explore workflow

About Krea 2 Turbo

Krea 2 is Krea AI's open-weight image generation model, built from scratch for aesthetic diversity, photorealism and creative control. It is deployable on HexGrid.cloud with dedicated GPU infrastructure, configurable inference workflows, custom checkpoints and LoRA support.

The open-weight release includes two variants: Krea 2 RAW and Krea 2 Turbo. RAW is the undistilled foundation checkpoint designed for flexible generation, fine-tuning and LoRA training. Turbo is an optimized 8-step distilled checkpoint designed for fast, high-quality text-to-image generation.

Krea 2 uses a 12-billion-parameter Diffusion Transformer architecture with a Qwen3-VL 4B text encoder, grouped-query attention and sigmoid-gated attention. RAW is trained for approximately 1K image generation, while Turbo supports output up to 2K resolution. The model specializes in expressive aesthetics, diverse artistic styles, cinematic photography and customizable visual generation.

What people use it for

Cinematic text-to-image

Generate photorealistic images directly from detailed text prompts, controlling subjects, environments, lighting, camera angles, depth of field, textures and cinematic composition. Create editorial photographs, cinematic stills and atmospheric visual concepts.

Artistic and stylized generation

Explore diverse artistic directions, including illustrations, anime, paintings, analog photography, risograph prints, collages and experimental graphic styles. Generate expressive images without being restricted to a single polished visual aesthetic.

Advertising and product visuals

Create campaign concepts, editorial product photography, promotional artwork and branded imagery. Experiment with lighting, product placement, backgrounds, materials and visual treatments for marketing and creative production.

Architecture and design visualization

Generate architectural concepts, interior designs, material studies, environmental visualizations and mood-driven design explorations. Compare lighting conditions, textures, compositions and aesthetic directions through natural-language prompts.

Custom styles and LoRA training

Fine-tune the RAW checkpoint on specialized visual datasets to create reusable characters, branded aesthetics, artistic styles or product identities. Apply compatible LoRAs to the Turbo checkpoint for accelerated inference.

Rapid creative exploration

Generate multiple visual variations from a single concept using configurable random seeds, image dimensions and sampling parameters. Use the Turbo checkpoint for fast experimentation with compositions, artistic directions and creative concepts.

Strengths

  • Native text-to-image generation with approximately 1K resolution on RAW and up to 2K resolution on Turbo
  • 12B-parameter Diffusion Transformer architecture trained from scratch for aesthetic diversity and creative exploration
  • Two open-weight checkpoints: RAW for fine-tuning and flexible sampling, Turbo for accelerated 8-step generation
  • Turbo supports high-quality image generation in just 8 denoising steps with classifier-free guidance disabled
  • RAW supports full diffusion sampling with configurable classifier-free guidance for detailed control over generation
  • Strong aesthetic diversity across photorealism, illustration, painting, anime, photography and experimental artistic styles
  • Cinematic visual quality with detailed control over lighting, composition, camera perspective, depth of field and atmospheric effects
  • Expressive rendering of analog photography, film grain, motion blur, soft focus and other deliberately imperfect visual aesthetics
  • Qwen3-VL 4B text encoder for natural-language understanding and detailed prompt conditioning
  • Grouped-query attention and sigmoid-gated attention for efficient transformer computation
  • SwiGLU feed-forward layers and zero-centered RMS normalization for improved transformer performance
  • Multilayer text-feature aggregation for combining coarse and fine semantic information from the text encoder
  • 3D axial rotary positional embeddings for spatial relationships within image generation
  • Rectified-flow diffusion training with resolution-aware timestep shifting
  • Support for square, landscape, portrait and custom image dimensions
  • Configurable denoising steps, guidance scales, timestep shifts, image dimensions and random seeds
  • Batch image generation with deterministic seed progression for reproducible variations
  • LoRA fine-tuning workflow designed around training on RAW and applying compatible adapters to Turbo
  • Optional LLM-based prompt expansion for enriching short descriptions into detailed visual instructions
  • BF16 inference with support for additional optimization techniques through compatible runtimes
  • Open-weight architecture supporting custom checkpoints, LoRA training and private GPU deployment

Limitations

  • The RAW checkpoint is trained for approximately 1K resolution. Higher-resolution generation is better suited to Turbo, which supports images up to approximately 2K.
  • The official open-weight release provides text-to-image generation only. Native image-to-image editing, inpainting and multi-reference composition are not exposed by the standard RAW and Turbo inference pipelines.
  • Style-reference conditioning, moodboard blending and creativity presets offered by Krea's hosted product are not included as standard inputs in the released RAW and Turbo checkpoints.
  • Turbo is optimized for 8 denoising steps with classifier-free guidance disabled. Its accelerated sampling configuration is not interchangeable with RAW's full sampling settings.
  • RAW typically requires approximately 52 inference steps with guidance scale 3.5 for recommended generation quality, resulting in longer inference times than Turbo.
  • The model generates static images only. Native video generation, animation, audio synthesis and synchronized audiovisual output are not supported.
  • Complex prompts involving exact object placement, precise spatial relationships or many interacting subjects may produce inconsistent compositions.
  • Although text rendering is supported, long text passages, small typography and densely formatted layouts may contain spelling or layout inaccuracies.
  • Exact reproduction of a specific visual style or character identity across many generations generally requires custom LoRA training or additional conditioning workflows.
  • Native 2K generation requires substantial GPU memory, particularly when using full-precision checkpoints or generating multiple images simultaneously.
  • Custom dimensions outside the model's recommended training resolutions may introduce composition artifacts, duplicated objects or reduced visual consistency.
  • High-resolution generation and increased sampling-step counts raise computational requirements and inference latency.
  • The official Krea 2 Community License permits commercial use only under specified conditions, including a company-wide annual revenue threshold below $1 million. Larger commercial deployments require a separate enterprise license.
  • The Krea 2 Community License also requires appropriate content filtering and compliance with applicable attribution, distribution and acceptable-use conditions.

Specifications

Generation, input and serving details.

ProviderKrea AIFamilyKrea-2CategoryimageWorkflows2

Generation

Modes
Text-to-image
Variants
Krea 2 RAW, Krea 2 Turbo
Architecture
Single-stream Diffusion Transformer with rectified-flow generation
Model parameters
12 billion
Transformer layers
28 diffusion transformer layers
Hidden dimension
6144
Attention heads
48 query heads, 12 key-value heads
Text encoder
Qwen3-VL 4B Instruct
Text conditioning
Multilayer feature aggregation with additional bidirectional text processing
Resolution
RAW: approximately 1K; Turbo: approximately 1K–2K
Default resolution
1024×1024 in the official inference script
Maximum recommended resolution
Turbo: up to approximately 2048×2048
Aspect ratios
Configurable square, landscape, portrait and custom aspect ratios
Image dimensions
Configurable width and height; dimensions are padded to multiples of 16
Output formats
PNG in the reference implementation; additional formats available through image post-processing
Inference steps
RAW: 52 recommended; Turbo: 8 recommended
Guidance scale
RAW: 3.5 recommended; Turbo: 0.0 recommended
Generation framework
Rectified-flow diffusion with velocity prediction
Sampling
Resolution-aware shifted timestep sampling
Attention mechanism
Grouped-query attention with gated sigmoid attention
Feed-forward network
SwiGLU
Normalization
Zero-centered RMSNorm with QK normalization
Positional encoding
3D axial rotary positional embeddings
Image decoder
Qwen Image VAE
Batch generation
Configurable number of images per generation request

Inputs and Controls

Text prompt
Required natural-language description of the target image
Input image
Not supported by the standard open-weight text-to-image pipeline
Reference images
Not exposed as native conditioning inputs in the standard RAW and Turbo inference pipelines
Image editing
Not natively supported by the standard open-weight checkpoints
Style control
Natural-language descriptions of artistic style, lighting, materials, composition and visual aesthetics
Camera and composition
Prompt-based control over perspective, framing, focal length, depth of field, lighting and photographic style
Negative prompt
No dedicated negative-prompt argument in the official reference CLI; CFG behavior is configurable
Inference steps
Configurable; recommended 52 for RAW and 8 for Turbo
Guidance scale
Configurable classifier-free guidance; 3.5 recommended for RAW and 0.0 for Turbo
Timestep shift
Resolution-dependent shift with configurable y1, y2 and optional constant mu
Minimum-resolution shift
y1: 0.5 by default
Maximum-resolution shift
y2: 1.15 by default
Turbo timestep shift
Constant mu of 1.15 recommended
Random seed
Configurable integer seed; default 0 in the official inference script
Seed progression
For multiple images, image i uses the base seed plus i
Image width
Configurable; default 1024 pixels
Image height
Configurable; default 1024 pixels
Number of images
Configurable; default 1
Checkpoint selection
RAW or Turbo
Prompt expansion
Optional LLM-based prompt enhancement using the published Krea 2 prompting guidance
Text rendering
Supported through natural-language prompts; quoted text is recommended for words intended to appear in images
LoRA training
Recommended on the RAW checkpoint
LoRA inference
Compatible RAW-trained LoRAs can be applied to Turbo through supported inference workflows
Custom checkpoints
Supported through compatible inference and fine-tuning implementations
Model precision
BF16 in the official reference implementation
Quantization
Additional optimized precision configurations depend on the inference runtime
GPU memory optimization
Configurable through compatible memory-management and optimized inference implementations
Image post-processing
Optional resizing, upscaling, format conversion and additional processing through custom workflows

Serving

Provider
Krea AI
API platform
HexGrid.cloud
Deployment
Private GPU deployment with configurable image generation and inference workflows
Generation workflows
Text-to-image generation, high-resolution image synthesis, batch generation and LoRA-conditioned inference
Model checkpoints
Krea-2-Raw and Krea-2-Turbo
Checkpoint formats
Safetensors
Output format
PNG images; additional formats through compatible output processing
Model availability
Open-weight checkpoints available for private deployment under the Krea 2 Community License
Model license
Krea 2 Community License Agreement
Commercial use
Permitted under the Community License for organizations with less than $1 million in total annual revenue, subject to its conditions; separate enterprise licensing required at or above the threshold
Fine-tuning
LoRA training and custom model adaptation supported, with RAW recommended as the training checkpoint
Model architecture
12B single-stream Diffusion Transformer with Qwen3-VL 4B text conditioning
Inference frameworks
Official Krea inference implementation, Diffusers, ComfyUI and compatible optimized inference runtimes
Infrastructure
Dedicated GPU infrastructure with private model execution and configurable inference pipelines
GPU requirements
High-memory GPUs recommended for full BF16 inference; exact requirements depend on checkpoint, resolution, batch size and memory optimizations
Memory optimization
Reduced-precision inference, model offloading and optimized attention implementations through compatible runtimes
Customization
Custom checkpoints, LoRAs, inference parameters, sampling schedules, prompt processing and image post-processing workflows
Performance
Turbo uses 8 denoising steps for accelerated inference; RAW uses approximately 52 recommended steps for full-quality sampling. Actual throughput depends on GPU hardware, resolution, batch size and optimization settings
Content safety
Deployments must implement appropriate content-filtering measures and comply with the Krea 2 Community License and Acceptable Use Policy

Explore related models