English (main description)

Simple, reliable image-to-image workflow for Krea 2 Turbo — since there are very few i2i workflows for Krea 2 yet, I'm sharing mine.

Two versions included

1. krea2_turbo_img2img.json — basic i2i: Load Image → auto-resize to 2MP (Krea 2's 2K limit) → VAE Encode → KSampler → output

2. krea2_turbo_img2img_facekeep.json — same pipeline plus two face-preservation systems:

   • Face protection mask: right-click the Load Image node → Open in MaskEditor → paint the face → the painted area is excluded from regeneration (original pixels preserved 100%). Mask is inverted internally, so just paint what you want to KEEP.

   • ReActor face swap (optional): swaps the original face back onto the generated image. Bypass the node (Ctrl+B) if you don't need it.

Required models (official Comfy-Org releases)

Place in the usual folders:

• diffusion_models/krea2_turbo_fp8_scaled.safetensors

• text_encoders/qwen3vl_4b_fp8_scaled.safetensors

• vae/qwen_image_vae.safetensors

Download: https://huggingface.co/Comfy-Org/Krea-2

Requirements

• Recent ComfyUI (Krea 2 support required)

• For the facekeep version only: ComfyUI-ReActor (install via Manager). If ReActor's onnxruntime throws a RuntimeError on Windows, reinstall the CPU build: pip uninstall -y onnxruntime-gpu onnxruntime then pip install onnxruntime inside your ComfyUI venv.

Settings guide

• Sampler: euler / simple, 8 steps, CFG 1.0 (Turbo defaults)

• Denoise: 0.4 = stay close to the source, 0.55 = balanced, 0.7 = strong restyle

• Input is auto-resized to max 2MP (Krea 2 supports up to 2K). To disable, bypass the resize node.

• ReActor: set face_restore_model to codeformer.pth or none (the exact filename depends on your install)

Tips

• For a 100% identical face, use the protection mask (paint the face) — ReActor alone gives you a "similar" face, not an identical one, because inswapper works at 128px.

• Both face systems can be combined.

日本語(補足)

Krea 2 Turbo用のimage-to-imageワークフローです。Krea 2のi2iワークフローがまだ少ないので公開します。

• 基本版: 画像読み込み→2MP自動リサイズ→VAEエンコード→KSampler

• 顔保持版: 上記+「顔保護マスク」(MaskEditorで顔を塗ると、その部分は再生成されず元のまま)+「ReActor顔スワップ」(不要ならCtrl+Bでバイパス)

必要モデルは Comfy-Org/Krea-2 の公式3点(krea2_turbo_fp8_scaled / qwen3vl_4b_fp8_scaled / qwen_image_vae)。設定は8ステップ・CFG 1.0・euler/simple、denoiseは0.4(忠実)〜0.7(大変化)。

顔を完全に同じにしたい場合はマスク保護を使ってください(ReActorは「似た顔」まで)。