A minimal ComfyUI workflow for inpainting with Krea 2. Load an image, mask the part you want to change, type what you want there, and run. Only the masked area is regenerated, and the rest of the image stays as it is.

How it works

• Everything is packed into one subgraph. The controls are the image (with its mask), a prompt, a seed, and dropdowns for the diffusion model, text encoder and VAE. It outputs the finished image.

• The mask is a latent noise mask combined with Differential Diffusion, so changes stay inside the masked area and blend into the surroundings instead of leaving a hard edge.

• It runs 10 steps at cfg 1 with a zeroed negative prompt, which suits Turbo’s fast distilled sampling.

• It uses only stock ComfyUI nodes, with no custom node packs required.

How to use it

1. Load your source image.

2. Right-click the image and choose Open in MaskEditor.

3. Use the Mask tool (not the Paint tool) to mark the area you want to change, then save. Leave it blank to change the whole image.

4. Type what you want changed in the prompt box on the Krea 2 Inpaint node.

5. Run it. The seed changes each run, so run again for a new result.

Tip: describe only the change, not the whole image. For example, write “a red leather jacket” instead of describing the entire scene.

Requirements

• A Krea 2 Turbo diffusion model, selected in the unet_name dropdown on the node (I built it with the INT8 version, but any Krea 2 Turbo file should work)

• The Qwen3-VL 4B text encoder (qwen3vl_4b_fp8_scaled), selected in the clip_name dropdown. The type is set to krea2 inside the subgraph.

• Qwen Image VAE (qwen_image_vae)