https://github.com/axu2/image-quilting

A numpy implementation of the paper "Image Quilting for Texture Synthesis and Transfer" (2001)
https://github.com/axu2/image-quilting

image-based-rendering texture-mapping texture-synthesis

Last synced: 2 months ago
JSON representation

A numpy implementation of the paper "Image Quilting for Texture Synthesis and Transfer" (2001)

• Host: GitHub
• URL: https://github.com/axu2/image-quilting
• Owner: axu2
• Created: 2020-01-09T08:11:53.000Z (over 5 years ago)
• Default Branch: master
• Last Pushed: 2024-02-16T07:33:03.000Z (over 1 year ago)
• Last Synced: 2025-03-27T21:52:04.980Z (3 months ago)
• Topics: image-based-rendering, texture-mapping, texture-synthesis
• Language: Jupyter Notebook
• Homepage:
• Size: 69.4 MB
• Stars: 19
• Watchers: 3
• Forks: 6
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# Image Quilting for Texture Synthesis and Transfer

![Demo](abstract_screenshot.PNG)

https://www.youtube.com/watch?v=QMiCNJofJUk

Image quilting is a technique of generating new images 

by stitching together patches of existing images.

It has applications of 

1) Texture synthesis, generating arbitrarily large textures from small real-world samples and 

2) Texture transfer, re-rendering an image in the style of another.

>The method

works directly on the images and does not require 3D information.

For more information, consult the original paper at https://people.eecs.berkeley.edu/~efros/research/quilting.html

All images in this readme come from the original paper or presentation at SIGGRAPH '01.

In this repository, we will be implementing the paper using Python and NumPy.

## Texture Synthesis

The algorithm starts with an input image and a block size:

![input block](input.png)

We then define a minimum cost path between the overlap of two blocks:



We then build up a synthesized image by tiling small blocks of the input image.

![build](build.png)

(a) Here, we just randomly choose blocks

(b) Here we pick blocks that have the least overlap error

(c) We do everything in (b) but also cut along the minimum error boundary.

## Texture Transfer

>[Here] just add another constraint when sampling: similarity to underlying image at that spot

![half](half.png)