Skip to main content
Inpainting allows you to selectively edit or regenerate specific regions of an image while preserving the rest. This is essential for fixing details, removing objects, or seamlessly adding new elements.

​
How Inpainting Works

  1. Define area: Create a mask marking the region to regenerate
  2. Encode: Convert image and mask to latent space
  3. Denoise: Only regenerate the masked area
  4. Decode: Blend seamlessly with unmasked regions
ComfyUI supports two inpainting approaches:
  • Standard models with VAEEncodeForInpaint
  • Dedicated inpaint models with InpaintModelConditioning

​
Standard Model Inpainting

Works with any Stable Diffusion model.
1
Load and prepare image
2
  • Add LoadImage node
  • Upload your image
  • Create a mask:
    • Paint white on areas to regenerate
    • Black = preserve
    • White = repaint
    • 3
    You can also use LoadImageMask to load a pre-made mask from the alpha channel.
    4
    Encode for inpainting
    5
  • Add VAEEncodeForInpaint node
  • Connect:
    • pixels: from LoadImage
    • vae: from CheckpointLoaderSimple
    • mask: from LoadImage mask output (or LoadImageMask)
  • Set grow_mask_by: 6 (default)
    • Expands mask slightly for seamless blending
    • Higher values = softer edges
    • Range: 0-64 pixels
    • 6
    Set up prompts
    7
  • Add CheckpointLoaderSimple
  • Add two CLIPTextEncode nodes
    • 8
    Prompt strategy:
    9
  • Describe what should be in the masked area
  • Include context from the surrounding image
  • Be specific about blending and style
    • 10
    Example (removing an object):
    11
    Positive: "grass lawn, natural lighting, photorealistic, seamless"
Negative: "object, person, artifacts, visible seams, inconsistent lighting"
    12
    Sample and decode
    13
  • Add KSampler:
    • latent_image: from VAEEncodeForInpaint
    • denoise: 1.0 (full regeneration of masked area)
    • steps: 25-40 (higher for better blending)
    • cfg: 7-8
  • Add VAEDecode and SaveImage
    • 14
    Generate
    15
    Press Ctrl+Enter. The masked area will be regenerated while preserving the rest.

    ​
    Workflow JSON: Standard Inpainting

    {
  "load_image": {
    "class_type": "LoadImage",
    "inputs": {
      "image": "photo.png"
    }
  },
  "checkpoint": {
    "class_type": "CheckpointLoaderSimple",
    "inputs": {
      "ckpt_name": "v1-5-pruned-emaonly.safetensors"
    }
  },
  "vae_encode_inpaint": {
    "class_type": "VAEEncodeForInpaint",
    "inputs": {
      "pixels": ["load_image", 0],
      "vae": ["checkpoint", 2],
      "mask": ["load_image", 1],
      "grow_mask_by": 6
    }
  },
  "positive": {
    "class_type": "CLIPTextEncode",
    "inputs": {
      "clip": ["checkpoint", 1],
      "text": "natural background, seamless blend, photorealistic"
    }
  },
  "negative": {
    "class_type": "CLIPTextEncode",
    "inputs": {
      "clip": ["checkpoint", 1],
      "text": "artifacts, seams, inconsistent, blurry"
    }
  },
  "sampler": {
    "class_type": "KSampler",
    "inputs": {
      "model": ["checkpoint", 0],
      "positive": ["positive", 0],
      "negative": ["negative", 0],
      "latent_image": ["vae_encode_inpaint", 0],
      "seed": 42,
      "steps": 30,
      "cfg": 7.5,
      "sampler_name": "dpm++ 2m karras",
      "scheduler": "karras",
      "denoise": 1.0
    }
  },
  "vae_decode": {
    "class_type": "VAEDecode",
    "inputs": {
      "samples": ["sampler", 0],
      "vae": ["checkpoint", 2]
    }
  },
  "save": {
    "class_type": "SaveImage",
    "inputs": {
      "images": ["vae_decode", 0],
      "filename_prefix": "inpaint"
    }
  }
}

    ​
    Inpaint Model Workflow

    Dedicated inpaint models (like SD 1.5 inpaint) produce better results for complex edits.
    1
    Load inpaint checkpoint
    2
    Use CheckpointLoaderSimple with an inpaint-specific model:
    3
  • sd-v1-5-inpainting.ckpt
  • sd-v2-inpainting.ckpt
    • 4
    Prepare conditioning
    5
  • Add InpaintModelConditioning node
  • Connect:
    • positive: from positive CLIPTextEncode
    • negative: from negative CLIPTextEncode
    • vae: from checkpoint
    • pixels: from LoadImage
    • mask: from LoadImage mask output
  • Set noise_mask: True (recommended)
    • Limits sampling to masked area
    • Improves quality and speed
    • 6
    This node outputs:
    7
  • Modified positive conditioning
  • Modified negative conditioning
  • Latent with noise mask
    • 8
    Sample with inpaint model
    9
  • Add KSampler
  • Use conditioning and latent from InpaintModelConditioning
  • Settings:
    • denoise: 1.0
    • steps: 30-50 (inpaint models benefit from more steps)
    • cfg: 7-9

    • ​
    Workflow JSON: Inpaint Model

    {
  "load_image": {
    "class_type": "LoadImage",
    "inputs": {
      "image": "photo.png"
    }
  },
  "checkpoint": {
    "class_type": "CheckpointLoaderSimple",
    "inputs": {
      "ckpt_name": "sd-v1-5-inpainting.ckpt"
    }
  },
  "positive": {
    "class_type": "CLIPTextEncode",
    "inputs": {
      "clip": ["checkpoint", 1],
      "text": "beautiful garden, colorful flowers, natural daylight"
    }
  },
  "negative": {
    "class_type": "CLIPTextEncode",
    "inputs": {
      "clip": ["checkpoint", 1],
      "text": "blurry, distorted, artifacts, low quality"
    }
  },
  "inpaint_conditioning": {
    "class_type": "InpaintModelConditioning",
    "inputs": {
      "positive": ["positive", 0],
      "negative": ["negative", 0],
      "vae": ["checkpoint", 2],
      "pixels": ["load_image", 0],
      "mask": ["load_image", 1],
      "noise_mask": true
    }
  },
  "sampler": {
    "class_type": "KSampler",
    "inputs": {
      "model": ["checkpoint", 0],
      "positive": ["inpaint_conditioning", 0],
      "negative": ["inpaint_conditioning", 1],
      "latent_image": ["inpaint_conditioning", 2],
      "seed": 123,
      "steps": 40,
      "cfg": 8.0,
      "sampler_name": "dpm++ 2m karras",
      "scheduler": "karras",
      "denoise": 1.0
    }
  },
  "vae_decode": {
    "class_type": "VAEDecode",
    "inputs": {
      "samples": ["sampler", 0],
      "vae": ["checkpoint", 2]
    }
  },
  "save": {
    "class_type": "SaveImage",
    "inputs": {
      "images": ["vae_decode", 0],
      "filename_prefix": "inpaint_model"
    }
  }
}

    ​
    Common Use Cases

    ​
    Object Removal

    Goal: Remove unwanted elements
    1. Mask the object to remove
    2. Prompt what should be there instead
    3. Use high grow_mask_by (8-16) for better blending
    Settings: 
    Prompt: "[describe background], natural, seamless"
Steps: 35-50
CFG: 7-8
grow_mask_by: 12

    ​
    Adding Elements

    Goal: Insert new objects or details
    1. Mask where new element should appear
    2. Detailed prompt describing the addition
    3. Include lighting/style context
    Example: 
    Prompt: "red sports car parked on street, realistic lighting matching scene, 
sharp focus, detailed"
CFG: 8-10 (higher for specific objects)

    ​
    Face/Detail Fixing

    Goal: Improve specific details
    1. Mask problem area tightly
    2. Describe desired result
    3. Low grow_mask_by (2-4)
    Settings: 
    Prompt: "beautiful detailed face, natural expression, sharp eyes"
Steps: 40-50
CFG: 7
grow_mask_by: 3

    ​
    Background Extension (Outpainting)

    Goal: Expand image borders
    1. Use ImagePadForOutpaint node to extend canvas
    2. Mask the new area
    3. Prompt continuation of the scene
    {
  "pad": {
    "class_type": "ImagePadForOutpaint",
    "inputs": {
      "image": ["load_image", 0],
      "left": 256,
      "top": 0,
      "right": 256,
      "bottom": 0,
      "feathering": 40
    }
  }
}
    The node outputs both padded image and automatic mask.

    ​
    Mask Creation Tips

    ​
    Manual Masking

    1. Use image editor (Photoshop, GIMP, etc.)
    2. White = regenerate, Black = keep
    3. Soft edges with blur for better blending
    4. Save with alpha channel

    ​
    Mask from Alpha Channel

    Use LoadImageMask: 
    {
  "mask": {
    "class_type": "LoadImageMask",
    "inputs": {
      "image": "mask.png",
      "channel": "alpha"
    }
  }
}
    Channels: alpha, red, green, blue

    ​
    Feathering

    grow_mask_by creates smooth transitions:
    • Small objects: 4-8
    • Medium areas: 8-16
    • Large regions: 16-32
    • Full image blend: 40+

    ​
    Advanced Techniques

    ​
    Multi-Region Inpainting

    1. Create mask with multiple separate areas
    2. Prompt describes all areas together
    3. Higher steps (40-60) for complexity

    ​
    Iterative Refinement

    1. First pass: Rough inpaint, denoise 1.0
    2. Save result
    3. Second pass: Load result, smaller mask, denoise 0.6-0.8
    4. Refine specific details

    ​
    Compositional Inpainting

    Use ConditioningSetMask for regional prompts: 
    {
  "masked_conditioning": {
    "class_type": "ConditioningSetMask",
    "inputs": {
      "conditioning": ["positive", 0],
      "mask": ["load_mask", 0],
      "strength": 1.0,
      "set_cond_area": "mask bounds"
    }
  }
}
    This applies prompt only to masked region, giving precise control.

    ​
    Parameter Guide

    ​
    grow_mask_by

    • 0: Hard edges, visible seams
    • 2-4: Tight blending, detail work
    • 6-8: Standard blending (default)
    • 12-16: Soft transitions, large areas
    • 32+: Very gradual blend, outpainting

    ​
    Steps

    • 20-30: Quick edits, simple fills
    • 30-40: Standard quality
    • 40-60: Complex inpainting, multiple objects
    • 60+: Challenging blends, high detail

    ​
    CFG Scale

    • 6-7: Natural blending, less artifacts
    • 7-8: Balanced (recommended)
    • 8-10: Strong prompt adherence, adding specific objects
    • 10+: Risk of oversaturation at mask edges

    ​
    Troubleshooting

    ​
    Visible seams

    • Increase grow_mask_by (try 12-20)
    • Lower CFG scale (try 6-7)
    • Increase steps (35-50)
    • Use softer mask edges

    ​
    Generated area doesn’t match

    • Describe surrounding context in prompt
    • Include lighting/style references
    • Try dedicated inpaint model
    • Increase CFG for stronger adherence

    ​
    Blurry results

    • Increase steps (40+)
    • Use better sampler (dpm++ sde karras)
    • Check VAE quality
    • Try dedicated inpaint model

    ​
    Changes bleed outside mask

    • Reduce grow_mask_by
    • Use hard-edged mask
    • Enable noise_mask (inpaint models)
    • Lower denoise slightly (0.95)

    ​
    Best Practices

    1. Match the style: Prompt should describe existing image style
    2. Contextual prompts: Reference surrounding elements
    3. Start simple: Test with small masks first
    4. Iterate: Run multiple generations, pick best
    5. Feather appropriately: Match grow_mask_by to edit size
    6. Consider dedicated models: Better for complex edits
    7. Higher steps: Inpainting benefits from extra refinement

    ​
    Next Steps

    • Combine with ControlNet for precise inpainting
    • Explore regional prompting with ConditioningSetMask
    • Try different inpaint-specific models
    • Experiment with outpainting for image extension

    Build docs developers (and LLMs) love

    Get started for freeTalk to us