AI Inpainting and Outpainting — How to Edit Specific Parts of Images Without Starting Over

Ever had a perfect photo ruined by a single detail — a lamppost growing out of someone's head, an unwanted reflection, or an object that shouldn't be there? Previously you either gave up (and reshot) or spent hours in Photoshop with clone stamp and masks. Now, AI lets you modify only the part that needs fixing, leaving everything else untouched. It's called inpainting (selective filling) and outpainting (extending beyond borders). This guide explains how they work, when to use them, and how to get professional results without the headache.

What is the difference between inpainting and outpainting?

Inpainting replaces or corrects a specific area within the image: a blemish, object, or person. The AI analyzes surrounding context (colors, textures, lighting) and generates matching pixels. The result looks like that detail never existed.

Outpainting expands the image beyond its edges — adding space to the right, left, top, or bottom. The AI imagines what might be outside the original frame and generates it plausibly. Useful for changing aspect ratios (e.g., vertical to square) or creating wider compositions.

We at Meteora Web use both techniques daily: removing brand tags from product photos, extending backgrounds for ad banners, or fixing small errors without reshooting. The time savings are huge — minutes instead of hours, and the final output is seamless.

Which tools support inpainting and outpainting?

Not all AI image generators handle these features equally. Here are the most used:

• Stable Diffusion (via Automatic1111, ComfyUI, or InvokeAI): most flexible. With the 'Inpainting' extension you mask the area and decide whether to regenerate it entirely or keep it close to the original. Free and open-source, but requires a decent GPU (or cloud like RunPod).
• Midjourney (v6+): supports inpainting via 'Vary (Region)' after generating an image. Select the area and describe what you want instead. Outpainting via 'Zoom Out' or 'Pan'. Convenient but subscription-based.
• Adobe Photoshop (Generative Fill): integrated in the 'Generative Fill' panel (AI). Works well for realistic photos and quick edits. Consumption-based credits (included in Creative Cloud subscription).
• DALL·E 3 (via ChatGPT Plus or Bing Image Creator): allows selecting an area and modifying it with a prompt. Very intuitive but less precise control: the AI may slightly expand or alter the context.

We recommend Stable Diffusion for those with technical skills wanting maximum control; Adobe Generative Fill for Photoshop workflows; Midjourney for those already in that ecosystem.

How to use inpainting to fix specific defects?

Consider a real scenario: a fashion product photo (a dress) with a crooked label and a small tear on the fabric. With inpainting, no need to reshoot — solved in a few clicks.

Step 1: Choose your tool and load the image

In Stable Diffusion (Automatic1111), go to 'img2img' tab and enable 'Inpaint'. Upload the image and use the brush to paint black over the area to remove (label and tear). The more precise the mask, the better the result.

Step 2: Configure parameters

Set 'Denoising strength' between 0.6 and 0.8: higher means more change. For local correction (replacing label with surrounding fabric), keep it around 0.5–0.6. Enable 'Inpaint at full resolution' to preserve details. Choose a stable sampler (e.g., Euler a) and medium mask area.

Step 3: Write the right prompt

The prompt for inpainting should describe what you want in place of the defect. Example: “smooth cotton fabric, no label, uniform texture, no visible seams”. If you write nothing, the AI fills the area with what it considers coherent (often works well). For specific contexts, guide it.

Step 4: Generate and iterate

Hit 'Generate'. If the result isn't perfect (e.g., an unnatural crease), repeat with a tighter mask or adjust denoising. Usually 2–3 attempts yield an invisible fix.

# Example API call with Stable Diffusion (via diffusers library)
from diffusers import StableDiffusionInpaintPipeline
import torch
from PIL import Image

pipe = StableDiffusionInpaintPipeline.from_pretrained(
    "runwayml/stable-diffusion-inpainting",
    torch_dtype=torch.float16
).to("cuda")

image = Image.open("dress.jpg")
mask = Image.open("mask.jpg")  # white mask on area to modify
prompt = "smooth cotton fabric, beige, no label"

result = pipe(prompt=prompt, image=image, mask_image=mask, height=512, width=512).images[0]
result.save("dress_fixed.jpg")

Common mistake: masking too little. If the area is small, the AI may lack context and generate artifacts. Slightly enlarge the mask.

Outpainting to extend the background — when does it make sense?

Outpainting is perfect when your starting image is too narrow for the required layout. Real examples: turning a 1:1 photo into a 16:9 banner for Google Ads, or creating a square Instagram image from a vertical portrait. Instead of reshooting with more space, the AI imagines the extension.

How to do it in Midjourney

Generate an image, then use 'Zoom Out 1.5x' or 'Zoom Out 2x' (or custom). Midjourney extends the image while keeping the style. You can also use 'Pan' to extend in a specific direction (e.g., to the right). Important: if the image has human subjects, the added limbs may be imperfect; better for backgrounds or landscapes.

How to do it in Stable Diffusion

Outpainting requires a mask covering the edges to extend. Upload the original image, create a white mask on the area you want to expand (e.g., the right half of the image). Set denoising to 0.7–0.9: higher for more creativity, lower to stay faithful to the existing background. The prompt should describe the extension content (e.g., “blue sky, green grass, mountains in the background”).

Economic benefit

We at Meteora Web estimated that outpainting reduces ad production time by 70% compared to reshooting or manual retouching. For an e-commerce client changing banner formats every campaign, it goes from 4 hours of work to 20 minutes. Considering the hourly rate of a graphic designer, the savings are clear.

What are the limitations of inpainting and outpainting?

It's not all magic. Common issues and how to avoid them:

• Unnatural results: AI may create repeating patterns or flat textures. Solution: smaller masks, lower denoising, detailed prompts.
• Loss of global coherence: when extending backgrounds, perspective may shift. Example: a floor line extends unrealistically. Use depth maps (in Stable Diffusion) to maintain perspective.
• Computational cost: cloud tools (Midjourney, Adobe) have per-usage costs. For repetitive work, a local GPU is cheaper.
• Human details: faces, hands, and fingers are the AI's worst enemies. If inpainting involves a face part, it often creates artifacts. Prefer fixing small defects, not reconstructing whole faces.

How much does AI inpainting/outpainting cost vs traditional retouching?

A professional retoucher using Photoshop might charge €30–50 per image (20–30 minutes work). With AI inpainting, the same job costs pennies in electricity and compute time. For a catalog of 100 products, the savings amount to thousands of euros.

We applied inpainting to a catalog of 500 clothing items: removed labels, fluff, and imperfections in 3 days versus the 3 weeks a human retoucher would have taken. The client spent 80% less and had the catalog ready for the season.

What to do now

If you want to start with inpainting right away, the fastest path is:

1. Download Stable Diffusion Web UI (Automatic1111) — free, portable version for Windows.
2. Prepare a test image with a small defect (e.g., an unwanted object in a landscape photo).
3. Follow the steps above: mask the area, set denoising 0.6, no specific prompt, generate.
4. Compare with the original. If unsatisfied, adjust mask or denoising.
5. Once you've mastered the flow, apply it to real projects: product photos, portraits, ads.

For outpainting, try taking a vertical image and extending it to square using Midjourney (if you have a subscription) or Stable Diffusion with a mask on the edges.

Remember: AI is a tool, not magic. Every output must be verified and refined. But once you learn the technique, modifying specific parts of images becomes a few-minute operation, not a nightmare.

For a deeper look into the whole AI image ecosystem, check our complete guide on AI for images covering all major tools.