Image Variations

lucasd · August 31, 2019, 8:03pm

Hello!, I’m trying to do the following;
Given an image, generate another image (or many) similar to the input image.
I looked at two techniques but they don’t work as I need:
Style Transfer: needs two images as input, I want to give one
Texture Synthesis: the given image is small, I want to use a 1000x1000 pixels image as input

What techniques are possible for what I want to do?

Thanks!

muellerzr · August 31, 2019, 8:09pm

Here’s one to get you started:

github.com

Nik2049/GAN_Image_generation/blob/master/WGAN-image-generation.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "%reload_ext autoreload\n",
    "%autoreload 2\n",
    "%matplotlib inline"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "from fastai.vision import *\n",

This file has been truncated. show original

A few things to note:
1000x1000 pixels image as an input will require a crazy amount of GPU memory and training time.

Think of GANs as make x like y or from y. In terms of how to set it up like that example, your y’s are whatever similar image you want to use. (The notebook I link above shows something like that)

Others may chime in to

lucasd · September 1, 2019, 12:48am

Hi!, thanks for the answer,

In my case I only have a y; only one and single input image, and I need to generate variations of the same size based only in that on input image.
I found this Github project which does something similar: https://github.com/mxgmn/ConvChain

But it confuses me, do I need Deep Learning, or just dive into Markov Chains?
(excuse me if the question is dumb!)

Thanks!