Krea2_Anything2RealCharacters V3 — ComfyUI Workflow Guide

This workflow is specially configured for Krea2_Anything2RealCharacters-V3, trained with KV-Cache technology using AI Toolkit. It allows you to transform any 2D, anime, or stylized artwork into photorealistic images with extreme consistency.

Please note: This image-to-image workflow operates differently from standard Text-to-Image setups. Follow the guide below for proper configuration.

📦 Required Models & Assets

1. LoRA Model:

   • Download: Krea2_Anything2RealCharacters - v3.0 | Civitai

   • Place the .safetensors file into your ComfyUI/models/loras/ directory.

2. Base Architecture (Model, CLIP, VAE):

   • Download: Comfy-Org/Krea-2 | Hugging Face

   • Download official Krea 2 weights and place them in their respective models/checkpoints (or diffusion_models), models/clip, and models/vae folders.

🧩 Essential Custom Nodes (Mandatory)

CRITICAL: Do NOT replace or omit these two nodes! They are essential for proper KV-cache feature alignment and image scaling.

• Krea 2 Ostris Edit Model Patch:

  • Repository: ostris/ComfyUI-Krea2-Ostris-Edit

  • Function: Handles the experimental AI Toolkit KV-cache edit patching for Krea 2.

• Easy_QwenEdit2509:

  • Repository: cardenluo/easy_qwenEdit_2509

  • Function: Resolves dimension offsets and image size limitations.

⚙️ Node Parameter Configurations & Recommendations

• Color Correct Node:

  • Tip: Testing shows that lowering the saturation slightly on the input image significantly boosts final realism. Adjusting pre-color balance allows you to control the overall photographic tone before image conversion.

• LayerUtility: ImageScaleByAspectRatio V2:

  • Tip: Set your maximum side resolution (Longest Edge) here. We recommend keeping it between 1024px and 2048px for optimal balance between speed and high-precision detail.

• Krea 2 Ostris Edit Model Patch:

  • Tip: Keep kv_cache strictly ENABLED/ON. Disabling KV-cache will result in a severe loss of image consistency, leading to blurry or broken outputs.