Krea 2 Turbo - Text to Image
ComfyUI in the Cloud
Deploy Krea 2 Turbo - Text to Image, from
$0.60 / hour
Krea-2 Turbo ComfyUI Workflow
A ComfyUI image generation workflow powered by Krea-2 Turbo, featuring INT8-quantized diffusion models, Qwen3-VL conditioning, LLM-based prompt enhancement, and optional LoRA support.
Workflow Overview
This workflow utilizes the Krea-2 Turbo model for AI image generation within ComfyUI. It combines an INT8-quantized diffusion model, Qwen3-VL 4B conditioning, a dedicated Qwen Image VAE, and optional LoRA integration for customized image styles.
The workflow also includes an LLM-based prompt enhancement system designed to refine or expand input prompts before image generation. Prompt enhancement is enabled with a maximum token configuration of 512, while LLM thinking mode is disabled.
INT8 quantization helps reduce the memory requirements of the diffusion model, while FP8-scaled conditioning provides a lower-precision representation for the Qwen3-VL text and vision encoder. The workflow supports external image dimension inputs and optional LoRA-based style customization.
Generation Settings
Prompt Enhancement
Enabled
LLM Max Tokens
512
LLM Thinking Mode
Disabled
LoRA
Disabled
Width
External Input
Height
External Input
LoRA Strength
0.80
Seed
45862206397178
Model Configuration
Image Generation Model (UNet)
INT8krea2_turbo_int8_convrot
The Krea-2 Turbo model serves as the primary image generation component responsible for the denoising process. This INT8-quantized version reduces model memory requirements compared with higher-precision weights. The Turbo model is designed for efficient image generation while supporting detailed text-based instructions.
Text / Vision Conditioning Model
Qwen3-VL 4Bqwen3vl_4b_fp8_scaled
The Qwen3-VL 4B model provides text and vision conditioning for the image generation process. It helps interpret input instructions and convert them into conditioning representations that guide the diffusion model. The FP8-scaled model configuration helps reduce memory requirements compared with higher-precision implementations.
Variational Autoencoder (VAE)
VAEqwen_image_vae
The Qwen Image VAE converts generated latent representations into visible images. During the decoding stage, the VAE transforms the latent output produced by the diffusion process into an image that can be previewed, saved, or processed further within ComfyUI.
Optional LoRA Model
Disabledkrea2_darkbrush
The workflow includes an optional Darkbrush LoRA configuration for modifying the visual style of generated images. The selected LoRA has a configured strength of 0.80 and an associated trigger phrase. LoRA application is currently disabled, meaning this optional style modification is not active in the displayed configuration.
Prompt Enhancement Configuration
The workflow includes a prompt enhancement feature that is enabled in the current configuration. This feature allows an LLM-based processing stage to refine or expand user-provided text prompts before they are used for image generation.
| Prompt Enhancement | Enabled |
| Maximum LLM Tokens | 512 |
| LLM Thinking Mode | Disabled |
The maximum token setting controls the configured token limit for the LLM processing stage. With thinking mode disabled, the workflow is configured to operate without the optional LLM thinking mode.
LoRA Configuration
The workflow supports optional LoRA integration for applying additional visual styles to generated images. A Darkbrush LoRA is selected in the configuration, although LoRA application is currently disabled.
| Enable LoRA | Disabled |
| LoRA Model | krea2_darkbrush |
| LoRA Strength | 0.80 |
| LoRA Trigger Word | muted minimalist sketch style |
Configured LoRA Trigger Phrase
muted minimalist sketch style
When enabled, the selected LoRA can influence the visual appearance of generated images. The configured trigger phrase provides a style instruction associated with the LoRA. Since LoRA is currently disabled, its configured strength and trigger phrase do not indicate that the style is actively being applied.
Additional Parameters
| Seed | 45862206397178 |
| Image Width | External Input |
| Image Height | External Input |
| Diffusion Model Precision | INT8 |
| Conditioning Model Precision | FP8 Scaled |
Surrounding Nodes and Workflow Integration
The Krea-2 Turbo subgraph consolidates model selection, prompt enhancement settings, optional LoRA configuration, and image generation parameters into a reusable workflow component. Width and height are exposed as external inputs, allowing other nodes to control the output dimensions.
In a typical ComfyUI image generation workflow, text prompts are supplied through upstream input nodes. With prompt enhancement enabled, the prompt can be processed through an LLM enhancement stage before being converted into conditioning information.
Qwen3-VL provides conditioning representations, while the Krea-2 Turbo diffusion model performs the image generation process. The optional LoRA component can modify the generation model when enabled, allowing additional style control.
The Qwen Image VAE decodes generated latent representations into visible images. These images can then be connected to preview, image-saving, or additional processing nodes. Exact upstream and downstream connections depend on the surrounding ComfyUI workflow.
Workflow Highlights
- ✓ Krea-2 Turbo: Uses the Krea-2 Turbo image generation model.
- ✓ Memory Efficiency: INT8 quantization helps reduce diffusion model memory requirements.
- ✓ Qwen3-VL Conditioning: Integrates the Qwen3-VL 4B FP8-scaled model.
- ✓ Prompt Enhancement: Enables LLM-based prompt processing with a configured maximum token value of 512.
- ✓ Optional LoRA Support: Includes a Darkbrush LoRA configuration with a strength of 0.80, currently disabled.
- ✓ Flexible Image Dimensions: Width and height are controlled through external inputs.
- ✓ VAE Decoding: Uses the Qwen Image VAE to convert generated latent representations into images.
- ✓ Reusable Subgraph: Organizes model selection, enhancement settings, and optional style controls in one component.
Workflow Summary
This ComfyUI workflow combines the Krea-2 Turbo INT8-quantized diffusion model with Qwen3-VL 4B FP8-scaled conditioning and the Qwen Image VAE. It features enabled LLM-based prompt enhancement with a maximum token setting of 512 and disabled LLM thinking mode.
The workflow also includes an optional Darkbrush LoRA configured at 0.80 strength with the trigger phrase "muted minimalist sketch style." LoRA is disabled in the displayed configuration.
Image dimensions are controlled through external width and height inputs. The subgraph organizes model components, prompt enhancement, and optional style customization for integration into larger ComfyUI image generation workflows.