Edit images using plain-English instructions, directly inside ComfyUI.

This workflow combines Krea 2 Turbo with Krea 2 Identity Edit v1.2 in an organized, fully local setup. Load an image, describe your edit, and generate. An optional second image can provide an additional visual reference.

Version 1.1 adds neural upscaling, adjustable final output sizing and a working-resolution cap, alongside an English installation guide and updated workflow notes.

What can you do?

• Change clothing, outfits and accessories.

• Adjust hair and eye colors.

• Change lighting, mood and time of day.

• Guide edits using an optional second reference image.

• Ask the model to preserve the subject, composition and surrounding scene.

• Upscale the edited result with a selectable neural upscale model.

Preservation depends on the instruction, settings and complexity of the edit. Exact identity and pixel-for-pixel preservation are not guaranteed.

Workflow features

• Clearly organized, color-coded sections.

• Editing instructions entered directly in English.

• Optional second image for visual prompt guidance.

• Automatic or manual working resolution.

• Adjustable generation-resolution cap, initially 1.5 MP.

• Adjustable reference fidelity.

• Optional Realism Engine LoRA, switched off by default.

• Neural upscaling with 2× final output scale by default.

• Local inference without a cloud image-generation API.

Quick start

1. Read the installation guide included in the ZIP.

2. Install the required custom nodes and download the model weights.

3. Open the workflow and select your local files in the model loaders.

4. Load your primary image. Enable the second image only when needed.

5. Choose your resolution and final upscale factor.

6. Write your editing instruction in English and queue the workflow.

Be specific about both the change and what should remain. For example:

“Change only the jacket to black leather. Preserve the subject’s face, pose, hairstyle, background and lighting.”

Start with 10 sampling steps and CFG 1, then adjust reference fidelity as needed.

Resolution and upscaling

The final scale is set to 2.0. Change it to 1.5 if you prefer a 1.5× output.

In AUTO mode, the workflow generates within the working-resolution cap, then restores the original input dimensions before upscaling.

In MANUAL mode, upscaling is applied relative to the generated working dimensions. With aspect-ratio preservation enabled, manual width and height act as maximum bounds.

Large AUTO inputs can still produce large final images, even with the generation cap enabled.

Installation and downloads

The ZIP includes:

• The English workflow JSON.

• An English installation guide in Markdown.

• The two small custom helper nodes used for resolution and final sizing.

Model weights are downloaded separately.

Required components include:

• Krea 2 Turbo, Qwen3-VL encoder and Qwen Image VAE.

• Krea 2 Identity Edit v1.2.

• comfyui-krea2edit.

• rgthree-comfy.

• A compatible neural upscale model, such as RealESRGAN_x2plus.

Saved loader paths come from my setup. Select your own downloaded files rather than assuming those paths exist on your machine.

If you already have the earlier Smart Edit helper package installed, check the guide before copying the included helpers to avoid duplicate node registrations.

Optional Realism Engine LoRA

Realism Engine is optional. Choose the Krea 2 variant, select it in Power LoRA Loader, and switch its entry on.

Leave the entry off if you do not want to use it. Keep the separate Identity Edit v1.2 loader enabled.

Hardware

Tested locally, 10 to 20 seconds per image on an RTX 5090, 64 GB system RAM (depends on the image resolution). This is the tested configuration, not a minimum requirement but I would recommend at least 10-12GB VRAM for a smooth Krea2 Turbo sampling, use more aggressive quantizations if needed (8GB VRAM). Memory use and generation time depend on resolution and model configuration.

Credits

Built using Krea 2 Turbo, Krea 2 Identity Edit, comfyui-krea2edit, rgthree-comfy and neural upscaling models from their respective creators. Optional LoRAs remain credited to their authors.

Feedback and example results are welcome!