ComfyUI Inpainting Guide: Edit Parts of an Image with AI

What is Inpainting?

Inpainting lets you modify a specific area of an image while keeping everything else unchanged. You paint a mask over the region you want to change, describe what should go there, and the AI regenerates only that area.

Common uses:

• Fix faces — correct expressions, eye color, or facial features
• Remove objects — erase unwanted items from a scene
• Swap clothing — change outfits on a character
• Add elements — insert new objects into an existing image
• Fix artifacts — repair AI generation errors in specific areas

Prerequisites

You need a checkpoint model and a source image to inpaint. No extra plugins are required — all necessary nodes are built into ComfyUI.

Building the Inpainting Workflow

Core Nodes

1. Load Checkpoint — loads your base model
2. Load Image — loads the image you want to edit
3. CLIP Text Encode (positive) — describes what you want in the masked area
4. CLIP Text Encode (negative) — describes what to avoid
5. VAE Encode — converts the image to latent space
6. Set Latent Noise Mask — applies the mask to define the editable region
7. KSampler — regenerates the masked area
8. VAE Decode → Save Image

Drawing the Mask

After loading your image in the Load Image node:

1. Right-click the Load Image node
2. Select Open in MaskEditor
3. Paint over the area you want to change — white areas will be regenerated, black areas stay untouched
4. Click Save to apply the mask

The Denoise Parameter

The most important inpainting parameter is denoise on the KSampler node:

Start low (0.3–0.5) and increase gradually until you get the result you want.

Step-by-Step Example: Changing Eye Color

1. Load a portrait image
2. Open the Mask Editor and paint over both eyes
3. Set the positive prompt to: bright green eyes, detailed iris
4. Set the negative prompt to: blurry, low quality, deformed
5. Set denoise to 0.4
6. Generate — the eyes change color while the rest of the face stays identical

Writing Good Inpainting Prompts

Unlike text-to-image, inpainting prompts should describe the masked region specifically, not the entire image:

Alternative Method: VAE Inpainting Encoder

ComfyUI also has a VAE Encode (for Inpainting) node that handles masking differently. Instead of using Set Latent Noise Mask, you connect the image and mask directly to this specialized encoder.

Both methods work. The Set Latent Noise Mask approach described above is simpler and more flexible for beginners. The VAE Inpainting Encoder can produce better edge blending in some cases, especially with dedicated inpainting models.

Common Issues and Fixes

Inpainted area doesn't blend with the rest of the image

• Lower the denoise value — high denoise generates content that looks disconnected
• Make the mask slightly larger to include some surrounding context
• Use a dedicated inpainting model (e.g., sd-v1-5-inpainting.ckpt)

Output changes areas outside the mask

• This usually means denoise is too high — lower it to 0.3–0.5
• Make sure the mask is properly saved in the Mask Editor

Inpainted area is blurry

• Increase steps to 25–30
• Raise cfg slightly (7–9)
• Add quality keywords to your prompt: sharp, detailed, high quality

Nothing changes at all

• denoise is too low — increase to at least 0.3
• Verify the mask exists — check that white areas are visible in the Mask Editor
• Make sure the mask is connected to the Set Latent Noise Mask node

Seam visible around the inpainted area

• Expand the mask to include more surrounding pixels
• Lower denoise to 0.3–0.4 for smoother blending
• Try the VAE Inpainting Encoder method instead

Related Guides

• Text to Image — Basic image generation
• Image to Image — Transform entire images
• Outpainting — Extend images beyond their borders