Saturday, May 6, 2017

This Month's Free Photos - In Your Dreams

Just a reminder: It's okay to leave comments!

This month is going to be a little different. I've been playing around with CNN. No, not that network—Google's convolutional neural network (aka CNN). Yes, they let you do that. For free. And it doesn't hurt a bit. Let me explain...

You may have noted that I do a little artwork. Very little, unfortunately. Still, I like to check out what's happening in the art world from the periphery on which I exist. Sometimes I even draw or paint something. A while back I stumbled onto a tool that really caught my attention and I've finally had a chance to work with it.

Google has created code for modifying images called DeepDream (or Deep Dream, depending upon the source). You can read about it HERE. I've spent a little time playing around with it on the web (link below) and thought I'd share a little of what I've done with you. Literally.

It works like this:

  1. You create a user id / password to get into the app (they don't ask for a lot of info)
  2. You upload an image (I've used both JPEG and PNG)
  3. Select whether you want to do a Deep Style or a Deep Dream (and how many levels!)
  4. Pick the parameters you want to apply
  5. Click on the generate button

The picture you make can be saved, downloaded, deleted, or re-edited, and can be kept private or made public.

This is a sample of what you can do with an image:

Original image of pocket watch
Original image of pocket watch

Both of the following images were made from the original using the Deep Style algorithms.

A Deep Style modification using the Simpsons themeA Deep Style modification

This one is done using one of the Deep Dream filters. You'll have to open it to a larger size to get the full effect.

A Deep Style alteration of the pocket watch image

Here's a closeup of the dial. Slightly disturbing, isn't it? Looks like the watch is the one doing the watching!

Disturbing closeup of the pocket watch dial

Now let's see what happens when we work with some nice flowers...

Original wood hyacinth photo I took last year
Original wood hyacinth photo I took last year

For the first step, here are the parameters:
  • Type = Valyrian
  • Resolution = normal 1200x1200
  • Inception depth = deep
  • Neural network layer = inception_3b-output

Hyacinth - first dream

For round 2:
  • Type = Neuron
  • Inception depth = deep
  • Transformation = none
  • Neural network layer = inception_3b-output 

Hyacinth - deeper dream

  • Type = Spirits
  • Inception depth = deep
  • Transformation = none
  • Neural network layer = inception_3b-output

Hyacinth - deeper yet

Even though a version of the original photo is here and I've given out the parameters I used, if you tried to do the same transformation the final output would not be exactly the same. There's something in the way the algorithms work that makes it slightly different each time.

Anyway, these are my free photos for the month! Use them any way you want!

And, as promised, here's a link to the site: Deep Dream Generator

If you're so inclined, you can read about convolutional neural networks HERE. Personally, I'd skip it and just look at the pretty pictures.

Check back again when you can. Maybe I'll have something even better!

Bonus image! click HERE to see an image generated from one of my photos  of a statue in McPherson Square, Washington, DC!

© 2017 K. R. Smith All rights reserved


  1. I would like to react with 'cool' and an exclamation mark. Wow! I bet it's fun to play with and the results are ... well, amazing. I love the pocket watch. The photos of it, too, of course, but the watch itself is quite beautiful

    1. Everybody seems to like the pocket watch pictures! I think I'll have to work on that some more!


Please feels free to post a comment!
Note: All comments will be moderated and will not be shown unless approved. Inappropriate comments will be removed.