Phase 2: Image Generation · 55 min · Python · Hugging Face Diffusers · Real-ESRGAN

Inpainting, Outpainting, Upscaling & Restoration

Generation is step one — post-processing (inpainting, upscaling, face restoration) is what makes images production-ready.

Hiring signal: Post-processing pipeline knowledge (inpainting, upscaling, face restoration) demonstrates you understand the full image generation pipeline, not just the generation step.

What you will learn

Implement inpainting: selective regeneration of specific image regions
Implement outpainting: extending image boundaries beyond original dimensions
Apply upscaling with ESRGAN and Real-ESRGAN for higher resolution
Use face restoration (CodeFormer, GFPGAN) to fix faces in generated images

The Problem

A team generates images with FLUX.1 Schnell at 1024x1024. The images look good, but:

A product image has a distorted hand — needs inpainting to fix
A hero image needs to be 4K for print — needs upscaling
A portrait has blurry faces — needs face restoration
A landscape needs to be wider — needs outpainting

Generation is only step one. Post-processing is what makes images production-ready. A pipeline without post-processing produces raw outputs that aren't suitable for production use.

What you'll build
A post-processing pipeline that takes a generated image and applies: (1) face restoration with CodeFormer, (2) 4x upscaling with Real-ESRGAN, (3) inpainting to fix a specific region. Measure processing time for each step.

Inpainting

Inpainting selectively regenerates specific regions of an image while keeping the rest unchanged.

from diffusers import StableDiffusionXLInpaintPipeline
from PIL import Image

# Load inpainting model
pipe = StableDiffusionXLInpaintPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16,
)

# Load image and mask
image = Image.open("generated_image.png").resize((1024, 1024))
mask = Image.open("mask.png").resize((1024, 1024))  # White = regenerate, black = keep

# Inpaint the masked region
result = pipe(
    prompt="a hand holding a coffee cup, detailed",
    image=image,
    mask_image=mask,
    num_inference_steps=30,
    guidance_scale=7.5,
    strength=0.8,  # How much to change the masked region
).images[0]

Inpainting Parameters

Parameter	Typical	Effect
`strength`	0.6-0.9	How much to change the masked region (1.0 = full regeneration)
`guidance_scale`	7.5	Prompt adherence in the masked region
`num_inference_steps`	25-30	Quality of inpainted region
`mask_blur`	4-8 pixels	Blends inpainted region with surrounding pixels

Creating Masks

import numpy as np
from PIL import Image, ImageDraw

# Create a mask (white = regenerate, black = keep)
mask = Image.new("L", (1024, 1024), 0)  # Black background
draw = ImageDraw.Draw(mask)

# Draw white rectangle over the region to fix
draw.rectangle([300, 400, 500, 600], fill=255)  # Fix this region

mask.save("mask.png")

Unlock the full lesson

You've read the first 2 sections. The rest of this lesson covers Outpainting, Upscaling with Real-ESRGAN, Face Restoration with CodeFormer, Production Post-Processing Pipeline, Pipeline Order Matters, Processing Time Expectations, Key Takeaways, What's Next — plus a hands-on lab, quiz, and project artifact.

Create a free account to unlock Phase 0 and Phase 1 of every course — no credit card.