Krea 2 Two-Image Masked Edit — True Latent Inpaint

A ComfyUI workflow for replacing a masked area in Photo 1 using the appearance or subject from Photo 2 as a reference.

Unlike workflows that generate the entire image and only overlay the result afterward, this workflow sends the mask into Set Latent Noise Mask → KSampler. Sampling is therefore restricted to the masked latent area, while the unmasked area is preserved from Photo 1.

A final feathered pixel composite is also applied to preserve the original image outside the mask and produce a cleaner boundary.

How to use

1. Load the base image into Photo 1.

2. Open the ComfyUI Mask Editor and paint the area you want to replace.

3. Load the reference image into Photo 2.

4. Edit the instruction prompt to describe what should be transferred from Photo 2.

5. Run the workflow.

White mask area = generated/edited area
Black mask area = preserved from Photo 1

Default settings

• Steps: 8

• CFG: 1

• Sampler: Euler

• Scheduler: Simple

• Krea 2 Turbo

• Krea 2 Identity Edit v1.2 LoRA

• Two-image grounded conditioning

• True latent noise mask

• Final feathered composite

Required custom node

• lbouaraba/comfyui-krea2edit

All other processing nodes used by the workflow are included with ComfyUI.

Required models

• krea2_turbo_fp8.safetensors

• krea2_identity_edit_v1_2.safetensors

• qwen3vl_4b_fp8_scaled.safetensors

• krea2RealVae_v10.safetensors

Model folder names may differ between installations. If a model is shown as missing, select the corresponding file from your local ComfyUI model folders.

Notes

• Photo 1 is the base image and supplies the mask.

• Photo 2 is used as a visual reference; it is not directly pasted over Photo 1.

• Expand the mask slightly when the edited area needs more space.

• If the boundary is visible, try increasing Grow Mask from 8 to approximately 16–24.

• The workflow JSON does not include any model weights or input images. Download the required resources separately and follow their respective licenses.