xDev Asia

Inpainting: chỉnh sửa vùng cụ thể trong ảnh. Outpainting: mở rộng ảnh ra ngoài. Instruction-based editing: InstructPix2Pix. Background removal và replacement. Hands-on với Stable Diffusion inpainting pipeline.

Generative AI: Tạo Hình ảnh & Video với AI

Phần 3: Thực hành Image Generation Nâng cao

xdev.asia

Giới thiệu

Inpainting cho phép chỉnh sửa vùng cụ thể trong ảnh bằng AI — thay đổi đối tượng, xóa watermark, hoặc thêm chi tiết mới. Outpainting mở rộng ảnh ra ngoài khung ban đầu. Đây là foundation cho 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]

Tạo 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 — Mở rộng ảnh

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 — Chỉnh sửa bằng Lệnh

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")

Tổng kết

Kỹ thuật	Mô tả	Use case
Inpainting	Edit vùng mask	Object replacement
Outpainting	Mở rộng ảnh	Extend composition
InstructPix2Pix	Edit bằng text	Style changes
Background removal	Tách foreground	Product photos
Object removal	Xóa đối tượng	Clean up images

📌 Bài tiếp theo: LoRA & Custom Model Training — tạo style riêng.

Bài 8: Inpainting, Outpainting & Image Editing với AI