3

Are there any open sourced algorithms that can take a couple of images as an input and generate a new, similar image based on that input? Or are there any resources where I can learn to create such an algorithm?

Mark
  • 131
  • 1
  • 1
  • 4

3 Answers3

2

I'm not an expert on that so you could probably get a better answer.

I'm not sure to understand what you're looking for. Are the couple of images about the same thing? Like pictures of cats and you want to generate a new cat based on these pictures? If that's what you want, you could probably take a look at Generative Adversarial Network (GAN) : Introduction. A GAN is made up of a Generator and a Discriminator. The goal of the discriminator is to distinguish the real data from the generated data. And the goal of the generator is to improve its generated data to look similar to real data. Then, if there are different cat images in your dataset, the generator will learn to create a new cat based on that dataset.

If what you're looking for is to take different images like a cat and a dog and generate a "catdog", you can take a look at Variational AutoEncoder (VAE). For example you can train two different VAE (Encoder/Decoder). One for cats, and one for dogs. Then you take the encoder of dogs and the decoder of cats. That what I saw one day, not sure if it really works.

Correct me if I'm wrong

Dlmss
  • 21
  • 4
  • Thank you, that sounds like a good starting point. Actually it is about similiar images. Specifically, I want to feed the algorithm with pictures of stellar nebula and I am hoping to create new pictures of nebula by the algorithm. It's for a game project that I am working on. – Mark May 22 '19 at 09:12
  • 1
    It probably exists some other methods that I don't know but GAN seems to be good here. To get a better idea, you can read this article ( https://towardsdatascience.com/image-generator-drawing-cartoons-with-generative-adversarial-networks-45e814ca9b6b) and replace Homer Simpson by stellar nebula. – Dlmss May 22 '19 at 09:27
1

This was exactly what I was looking for;-) will be following!

I do have something that might be of interest to you, an algorithm that generates pictures from textual input. Here is the link to the site where you can try it out:

https://hypnogram.xyz/

Having played with it a little it seems that the text is used to search online and the pictures found online are then merged together according to a rule set.
But that is just a wild guess. It also seems that the images generated by the algorithm are unique and even when you use the same string twice the generated picture does not repeat itself. But again, that is what it seems like to someone who does not actually know, I only recently stumbled on the site and have yet to read more about the process. Enjoy!

TJx
  • 11
  • 1
1

I'm not an expert and I'm aware this is an old question so you've probably found something of interest since.

For latecomers, an interesting lead could be ControlNET, an extension for Stable Diffusion. This extension includes various models to analyze an input image and control the output. Some of the models include detection of pose, depth, contours... It's extremely useful to create iterations of an existing image, which seems to be your objective.

curious
  • 111
  • 3