Lesson 8: Inpainting, Outpainting & Image Editing with AI

Generative AI: Create Images & Videos with AI

Part 3: Practicing Advanced Image Generation

xdev.asia

Introduction

Inpainting allows you to edit specific areas in photos using AI — change objects, remove watermarks, or add new details. Outpainting expands the image beyond the original frame. This is the foundation for AI-powered image editing.

---

1. Inpainting Pipeline

from diffusers import StableDiffusionXLInpaintPipeline
from PIL import Image
import torch

pipe = StableDiffusionXLInpaintPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16,
)
pipe.to("cuda")

# Original image + mask (white = area to edit)
image = Image.open("photo.png").resize((1024, 1024))
mask = Image.open("mask.png").resize((1024, 1024))

result = pipe(
    prompt="a golden retriever puppy",
    image=image,
    mask_image=mask,
    num_inference_steps=30,
    guidance_scale=7.5,
    strength=0.99,  # how much to change masked area
).images[0]

Create Mask programmatically

from PIL import Image, ImageDraw
import numpy as np

# Mask thủ công: rectangle vùng cần edit
mask = Image.new("L", (1024, 1024), 0)  # black = keep
draw = ImageDraw.Draw(mask)
draw.rectangle([300, 200, 700, 600], fill=255)  # white = edit

# Mask từ segmentation (SAM)
from segment_anything import SamPredictor, sam_model_registry

sam = sam_model_registry["vit_h"](checkpoint="sam_vit_h.pth")
predictor = SamPredictor(sam)
predictor.set_image(np.array(image))
masks, _, _ = predictor.predict(point_coords=np.array([[500, 400]]),
                                  point_labels=np.array([1]))
mask = Image.fromarray(masks[0].astype(np.uint8) * 255)

---

2. Outpainting — Expanding the image

def outpaint(pipe, image, direction="right", extend_px=512):
    """Mở rộng ảnh bằng inpainting"""
    w, h = image.size

    if direction == "right":
        canvas = Image.new("RGB", (w + extend_px, h), (0, 0, 0))
        canvas.paste(image, (0, 0))
        mask = Image.new("L", (w + extend_px, h), 0)
        draw = ImageDraw.Draw(mask)
        draw.rectangle([w - 50, 0, w + extend_px, h], fill=255)  # overlap 50px
    elif direction == "down":
        canvas = Image.new("RGB", (w, h + extend_px), (0, 0, 0))
        canvas.paste(image, (0, 0))
        mask = Image.new("L", (w, h + extend_px), 0)
        draw = ImageDraw.Draw(mask)
        draw.rectangle([0, h - 50, w, h + extend_px], fill=255)

    result = pipe(
        prompt="continue the scene naturally",
        image=canvas,
        mask_image=mask,
        num_inference_steps=30,
    ).images[0]

    return result

---

3. InstructPix2Pix — Editing with Commands

from diffusers import StableDiffusionInstructPix2PixPipeline

pipe = StableDiffusionInstructPix2PixPipeline.from_pretrained(
    "timbrooks/instruct-pix2pix",
    torch_dtype=torch.float16,
)
pipe.to("cuda")

image = Image.open("photo.jpg")

# Chỉnh sửa bằng text instruction
edited = pipe(
    prompt="make it a winter scene with snow",
    image=image,
    image_guidance_scale=1.5,  # how close to original
    guidance_scale=7.0,
    num_inference_steps=20,
).images[0]

# Ví dụ khác
# "turn the cat into a tiger"
# "make it sunset lighting"
# "add flowers to the garden"
# "change the shirt color to blue"

---

4. Background Removal & Replacement

from rembg import remove
from PIL import Image

# Remove background
input_img = Image.open("person.jpg")
output = remove(input_img)  # transparent background
output.save("person_nobg.png")

# Replace background with AI
def replace_background(person_img, bg_prompt, pipe):
    # 1. Remove background
    fg = remove(person_img)

    # 2. Generate new background
    bg = pipe(
        prompt=bg_prompt,
        num_inference_steps=25,
        width=person_img.width,
        height=person_img.height,
    ).images[0]

    # 3. Composite
    bg.paste(fg, (0, 0), fg)  # paste with alpha mask
    return bg

result = replace_background(
    Image.open("person.jpg"),
    "a tropical beach at sunset, palm trees",
    pipe
)

---

5. Object Removal

from diffusers import StableDiffusionXLInpaintPipeline
from segment_anything import SamPredictor

# 1. Select object to remove (click point)
predictor.set_image(np.array(image))
masks, _, _ = predictor.predict(
    point_coords=np.array([[x, y]]),  # click on object
    point_labels=np.array([1])
)

# 2. Dilate mask slightly
import cv2
kernel = np.ones((15, 15), np.uint8)
dilated = cv2.dilate(masks[0].astype(np.uint8), kernel) * 255
mask = Image.fromarray(dilated)

# 3. Inpaint to remove
result = pipe(
    prompt="empty background, natural scenery",
    image=image,
    mask_image=mask,
    num_inference_steps=30,
).images[0]

---

6. Batch Image Editing

import os
from pathlib import Path

def batch_edit(input_dir, output_dir, pipe, edit_prompt):
    """Batch edit images with InstructPix2Pix"""
    Path(output_dir).mkdir(exist_ok=True)

    for img_file in Path(input_dir).glob("*.{jpg,png}"):
        image = Image.open(img_file).resize((512, 512))
        edited = pipe(
            prompt=edit_prompt,
            image=image,
            num_inference_steps=20,
        ).images[0]
        edited.save(Path(output_dir) / img_file.name)
        print(f"✓ {img_file.name}")

batch_edit("photos/", "edited/", pipe, "make it a pencil sketch")

---

Summary

Engineering	Description	Use cases
Inpainting	Edit mask area	Object replacement
Outpainting	Expand photo	Extend composition
InstructPix2Pix	Edit with text	Style changes
Background removal	Split foreground	Product photos
Object removal	Delete object	Clean up images

📌 Next article: LoRA & Custom Model Training — create your own style.