CGFlowStudioGUIDES
Back to All Tutorials
Intermediate22 min read

Complete Krea-2 ComfyUI Workflow Guide: Turbo T2I, RAW Training & Style LoRA Reference Setup

Step-by-step tutorial on running Krea AI's Krea-2 natively in ComfyUI — covering Krea 2 Turbo (8-step) vs RAW (52-step), Qwen3-VL-4B text encoding, style LoRAs, and style reference workflows.

By CGFlowStudio
•
2026-08-20

Complete Krea-2 ComfyUI Workflow Guide: Turbo T2I, RAW Training & Style LoRA Reference Setup

Krea-2 from Krea AI is a next-generation open-weights image foundation model built from scratch with a focus on high aesthetic quality, artistic versatility, and style control. Featuring native Qwen3-VL-4B multimodal text encoding, built-in prompt enhancement, official style LoRAs, and zero-shot Style Reference conditioning, Krea-2 gives creators unprecedented artistic freedom.

Krea-2 is available in two core editions: Krea 2 RAW (a 52-step base model engineered for fine-tuning and LoRA training) and Krea 2 Turbo (an 8-step distilled edition for fast, high-quality daily generation).

With official native support available in ComfyUI, setting up Krea-2 is fast and intuitive. In this step-by-step guide, we walk you through installing model files, loading official template workflows, configuring style LoRAs, and running both Text-to-Image Turbo and Style Reference generation pipelines.

1. Introduction & Core Capabilities

Before jumping into node setup, let’s look at what sets Krea-2 apart from traditional diffusion models:

💡
Seamless LoRA Compatibility Across RAW & Turbo: Custom style LoRAs trained on the full 52-step Krea 2 RAW base model transfer seamlessly to the 8-step Krea 2 Turbo model without quality loss or weight degradation.

RAW vs. Turbo Model Edition Comparison

Training Base 52 Steps (Full)

Krea 2 RAW

Full-precision 52-step base model. Highly malleable with high aesthetic ceiling; engineered specifically for custom fine-tuning and LoRA training.

Daily Generation 8 Steps (Distilled)

Krea 2 Turbo

Distilled 8-step model for rapid generation. Renders 1024x1024 to 2K imagery in under 3 seconds while supporting all RAW-trained LoRAs.

2. Getting Started – Loading Official Templates

ComfyUI provides built-in native workflow templates for Krea-2.

1 Update ComfyUI Core

Update ComfyUI to the latest version to register native Krea-2 loader nodes and subgraphs.

2 Open Template Library

Click Templates in the ComfyUI top menu and search for Krea-2 (or Krea-2: Text to Image).

3 Select Template

Choose from Text to Image (Krea-2 Turbo) or Style Reference (Krea-2).

3. Model Installation & Directory Structure

Download and place required Krea-2 checkpoints, Qwen3-VL-4B text encoders, VAEs, and style LoRAs into your local ComfyUI model subdirectories:

📂 Krea-2 Model Component Directory Map ComfyUI/models/
📂 ComfyUI/models/
📁 diffusion_models/
📄 krea2_turbo_fp8_scaled.safetensors (Turbo 8-Step Model)
📄 krea2_raw_fp8_scaled.safetensors (RAW 52-Step Model)
📄 krea2_style_ref_int8.safetensors (Dedicated Style Ref Model)
📁 text_encoders/
📄 qwen3vl_4b_fp8_scaled.safetensors (Qwen3-VL-4B Encoder)
📁 vae/
📄 qwen_image_vae.safetensors
📁 loras/
📄 krea2_darkbrush.safetensors
📄 krea2_softwatercolor.safetensors
📄 krea2_retroanime.safetensors
📄 krea2_style_reference_lora.safetensors

(Official models can be downloaded from Comfy-Org/Krea-2 on Hugging Face).

4. Workflow Walkthrough 1: Krea-2 Turbo Text-to-Image

The Turbo Text-to-Image workflow is designed for single-click daily generation.

Krea-2 Turbo Flow
Qwen3-VL-4B Text Encoder → Krea-2 Turbo Subgraph (8 Steps) → CustomCombo (Style LoRA) → SaveImage

Step-by-Step Instructions

  1. Load Template: Open Templates -> Image -> Krea-2: Text to Image.
  2. Enter Prompt: Type your prompt into the Text to Image (Krea-2 Turbo) subgraph input.
  3. Configure Prompt Enhancement: Toggle prompt_enhance = true to automatically expand short prompts with rich aesthetic modifiers.
  4. Set Target Resolution: Set target megapixels via ResolutionSelector (e.g., 1.0 MP = 1024x1024, 2.0 MP = 1408x1408).
  5. Run Queue: Click Queue Prompt. Generation completes in 8 sampling steps.

5. Official Style LoRAs Reference Guide

Krea AI provides a collection of official style LoRAs designed to inject instant artistic aesthetics into Krea-2 generations:

LoRA Name Style Trigger Word Recommended Strength Visual Aesthetic
krea2_darkbrush darkbrush style 0.7 – 1.0 Moody digital oil painting with heavy impasto stroke contrast.
krea2_softwatercolor soft watercolor 0.6 – 0.8 Delicate pastel watercolor wash with natural paper bleeding.
krea2_retroanime retro anime 90s 0.8 – 1.0 Classic 90s hand-drawn cel anime with warm chromatic aberration.
krea2_vintagetarot vintage tarot card 0.7 – 0.9 Intricate linework, gold foil accents, and woodcut engraving style.

How to Enable a Style LoRA

  1. Select your target LoRA in the CustomCombo node dropdown.
  2. Toggle enable_lora? = true inside the Krea-2 subgraph.
  3. Adjust lora_strength (0.7 is recommended to balance style against prompt accuracy).

6. Workflow Walkthrough 2: Krea-2 Style Reference

The Style Reference workflow generates brand-new images while locking the exact artistic style of 1 or 2 reference images.

Style Reference Flow
Load Image (Style Ref 1)
Load Image (Style Ref 2)
Krea2StyleReferenceLoader
[krea2_style_ref_int8]
KSampler → Styled Image Output

Step-by-Step Instructions

  1. Load Template: Open Templates -> Image -> Style Reference (Krea-2).
  2. Upload Reference Image(s): Load 1 or 2 reference images containing your desired color palette and art texture into Load Image nodes.
  3. Write New Subject Prompt: Describe the new subject you want to generate:
    • Example Prompt: "A futuristic Cyberpunk motorcycle parked in a wet alleyway." (The workflow applies the watercolor or oil paint style from your reference images automatically).
  4. Run Queue: Click Queue Prompt.

7. Prompting Tips & Best Practices

Krea-2 features a built-in Qwen3-VL prompt enhancer that expands short prompts automatically.

Good vs. Weak Prompt Comparison

❌ Weak Prompt

"A portrait of a knight."

Generic composition. Lacks lighting details, armor materials, and stylistic direction.

✅ Strong Prompt

"A moody portrait of a veteran knight in ornate silver armor, darkbrush style, dramatic rim lighting, atmospheric mist, cinematic 35mm composition."

Includes explicit subject descriptors, style trigger words (`darkbrush style`), and lighting specs.

8. Practical Tips & Troubleshooting

⚡ Selecting the Right Precision Variant

Use krea2_turbo_fp8_scaled.safetensors for standard 16GB GPUs. For 8GB–12GB VRAM GPUs, select krea2_turbo_int8.safetensors or MXFP8 quantized variants.

🎨 LoRA Trigger Words

Always include the corresponding trigger word (e.g., darkbrush style or soft watercolor) in your prompt text when enabling style LoRAs to activate optimal weight conditioning.

📂 Saved Image Location

All rendered images are saved directly to your ComfyUI/output/ directory with embedded workflow metadata.

9. Conclusion & Next Steps

Krea-2 combines ultra-fast 8-step Turbo generation with unparalleled style LoRA flexibility and zero-shot Style Reference conditioning. By starting with the Krea-2 Turbo Text-to-Image workflow and experimenting with official style LoRAs, you can create diverse artistic imagery directly on your local workstation.

Next Steps to Explore

  • Test combining Krea-2 Turbo with Style Reference nodes for custom artistic projects.
  • Train custom LoRAs using Krea 2 RAW (52-step base) and load them inside Krea 2 Turbo.
  • Explore more image guides on our ComfyHub Tutorials Directory.