Google DeepDream

I've gotten my hands on getting DeepDream to work on my laptop. There's no one single compiled app but after a bit of tinkering I was able to get it to work. I decided to feed some images of mine into the thing which results in pretty neat, convoluted images of what the AI sees and what we're supposed to see.

Here's a Buddhabrot.

Thanks very much for showing this, as I did not know nothing about this and seems tobe cool and interesting, although regretably this seems to be only for programmers or someone that knows how to use this

As this is not a normal executable file that you can install as it is from Github deepdream


As you are an expert in this, Could you be so kind to explain HOW can this be used, for other non expert people?

How to install this and be able to use it and feed the images? How does it work?

These Google "Deep Dream" images are weirdly mesmerising

Wired magazine WEIRD results from Google deepdream

I wonder HOW these amazing images can be done using this new DeepDream?

Now You Can Turn Your Photos Into Computerized Nightmares With 'Deep Dream

Deep Dream article

---------------------------------------------

Philip K. Dick's made the "Do Androids Dream of Electric Sheep?"

well now there is an answer

Yes, androids do dream of electric sheep

and many more in this google search

More!

Yeah... stuff Inside the dots is very interesting. Love it and also animals in your image.

Well the network is trained based on a large data set you can grab online. If you can grab hold of the variables google's seeing and calling in the image on the first iteration, you should be able to reverse engineer the thing.

This is just a guess btw.

Keyword: The Thing



Now all we need are king crab upside down man heads.

Explaining in some possible way how DeepDream works and what it does

Artificial Neural Networks Can Day Dream–Here's What They See

and the extended explanation in some way is found here

Inceptionism: Going Deeper into Neural Networks

Dockerized deepdream: Generate ConvNet Art in the Cloud - Brain-dead simple instructions for programmers

I have just found that seems to be a new alternative deepdream release that seems to be simpler although still only for programmers that understands it

Quote

Dockerized deepdream: Generate ConvNet Art in the Cloud github.com/VISIONAI/clouddream Google recently released the deepdream software package for generating images like which uses the Caffe Deep Learning Library and a cool iPython notebook example. Setting up Caffe, Python, and all of the required dependencies is not trivial if you haven't done it before! More importantly, a GPU isn't required if you're willing to wait a couple of seconds for the images to be generated. Let's make it brain-dead simple to launch your very own deepdreaming server (in the cloud, on an Ubuntu machine, Mac via Docker, and maybe even Windows if you try out Kitematic by Docker)! Motivation I decided to create a self-contained Caffe+GoogLeNet+Deepdream Docker image which has everything you need to generate your own deepdream art. In order to make the Docker image very portable, it uses the CPU version of Caffe and comes bundled with the GoogLeNet model. The compilation procedure was done on Docker Hub and for advanced users

Read the whole and complete text on the link

It may be brain-dead simple for programmers because I do not understand it and do not know what is being talked about, or I am brain dead?

Also it seems that all is command line based, and I personally do not like any software based on command line without any graphical interface, sorry that I am a visual person and do not feel right with words only based software that is only for programmers and coders

Really cool fractal style DeepDream video

Quote
Here's What Google's Trippy Deep Dream AI Does To A Video Selfie Totally mind-bending.

Great and awesome fractal style compositions on the video, really cool

Quote
Skybase Spaceray, the feature set doesn't change. What's different is that you can't change variables around. Which doesn't necessarily make better or worse images. So basically, it's the same thing.

Thanks, maybe the features doesn´t change, but if you can´t change the variables, I suposse that it may make a big difference, or not?

Is like in comparison, having filter forge filters that render one image with one preset and you can´t change any of the values?

As said, I think I will wait for an easier way done for non-programmers

TEST with the online tool

It puts "computer is now dreaming" and some sheeps jumping

It may take maybe a WEEK to complete?

Maybe this is because there is a huge list of people uploading images and it takes time to process them all

This is through a service that sends it on for you to Deep Dream. Can't run the app from my Mac, graphic card is Intel. The size is reduced at deepdreamit.com.

Maan.... lol finally it has come down to this.

Not that it's wrong, just hope $1.99 goes into the right causes.

Well this bothers me, so send me 1 image you wanna see google deep dream on and I can process it. It's gotta be relatively small so don't expect print quality out of this.


[UPDATE] After a bit of tinkering I was able to produce larger images although it takes significantly longer to work itself. You can try sending me reasonably large images, but not too large. I doubt it'd work with ridiculously huge resolutions. It's pretty RAM heavy as well.

DeepDream has reached Pinterest and Flickr (and many more places)

Pinterest googles deep dreams algorithm and inceptionism

One of the many Flickr album pages

DeepDream images collection Flickr by Kyle McDonald

Also seems that there is another website that processes DeepDream images

https://dreamdeeply.com/

I took the liberty of enlarging the following image so that I could look at it. It was flat as a pancake brightness & color wise, so I punched it up a bit with Vibrance. Poor Kiko, the scientific experiments, oh the agony (computed in a matter of weeks from deepdreamit.com)

It has been already a month since this appeared, has there been any news about it having a new tool or new GUI, or is it still the same?

I mean that maybe someone could have made anything new to be able to use in an easier way that may have appeared in some news

Here's a snapshot of DeepDreamer. Didn't know what I was doing, this is what came out. First it looks an awful lot like reaction diffusion. Second, it's not free as you can see from the partially crippled screen (they want $14.99US). Pretty cheeky for a beta product, we used to call this crippleware.

It honestly looks fine, it works perfectly for me. The first snapshot came out that way because your settings: Layer 3A 1x1 and your iteration count etc. DeepDream by default has various alternative settings accessible via class which you can load up to produce pretty intense images. Keep in mind that Deep Dream does seem to favor image with low contrast for more interesting results.

I think, under RealMac's philosophy for dev costs, charging $14.99 is technically fair, judging fr om the actual stability and speed of the product. However, it's a curiosity. I can run Deep Dream via Docker with a couple steps and it only takes a couple copy-pastas of code before I get other results. It's a bit slower, but I didn't pay for my results.

If anything I'd rather drop off money for Deep Styles which recently hit waves of interested people. That looks more entertaining to me than mashing up your pictures into psychedelic bonkers of dogs and weird caravans.

Ultimately I feel DeepDream is supposed to be an exploration into deep learning algorithms and that the image making function we see today is really a byproduct. It's open source for the reason that it opens people to ideas about the future wh ere machine learning can possibly make lives easier and more intersting.

Quote
As you mention, it seems that "eyeballs" and dogs and other objects show up too often, the technology appears to be somewhat limited at this point. Might be how the data is sliced and the limitations of the pattern recognition that is going on. It is amazing that it works at all!

Well what you really do is change the caffe model to other things, for example, there are a couple freely available online that recognizes places, flickr photographs. Again the intention of those models is for it to be able to recognize what humans see daily as well as places. You can swap out the default model for another relatively easily if you're working off the version that I'm using, but the RealMac version probably doesn't let you do that due to copyright restrictions of some of the models (i.e. Flickr trained model). Also the fact with the default model is that it probably has an absurd amount of dogs and other animals in it. So that's really a bias you're seeing.

Models are basically "trained" to recognize forms so you can make it recognize anything. For example you can use it for handwriting, daily objects, places, and you can of course get more specific with these. As an example, there's already a green-screen cutout method that utilizes deep-learning methods to specifically produce accurate greenscreen chroma keys.

Quote
Given that it is open source, I think I'll wait to see if someone comes up freeware for the Mac that doesn't require a specific GPU. For now open source is not a solution for most Mac users without the required hardware.

You also shouldn't need GPUs to run the images through. It's clearly faster but the process is relatively fast enough.

Very nice! If you have time, can you write up a rough installation guide? Otherwise don't worry, I'm probably just going to try it myself anyway.

Here's another interesting one: This one includes animation features.
https://github.com/mbartoli/neural-animation

Thanks very much Skybase for the Links you have put, good to know that there is an alternative even it is not free and cost 15$, but is worth to have a GUI and it work right, it may then worth to buy it, and is not a fortune to spend.

Thanks also very much to Rick Duim for the examples and comments

Neural-Style works pretty well for 512 pixel wide output, but it quite slow for anything larger than that (as mentioned above). I'm doing a 1280 pixel wide output image for 300 iterations now. It's up to 250, I'll post the results. Estimated time to complete: 22 hours.

This program needs a UNIX box with terabytes of memory for larger image output. It's amazing the Mac can handle the swapping in and out of memory. The algorithm is quite impressive given the research nature of it (MIT license). JC Johnson is actively updating the project LUA files, so it's a work in progress.

It seems that this is growing and more popular

Ok so it works. Not the best of qualities but it works. It's a mixture between my cat and a digital painting I did a couple years ago.

Although at this rate it's clearly faster running this on GPU. I think it's time for AmazonEC2. I still have a lot to read on but as far as my understanding goes, but the benefits are clearly there.

The thing is, I've been following a community of deep-dream enthusiasts and they've already setup an AMI to use. Otherwise considering to tinker with AmazonEC2 to make my own mini renderhub. The thing is a lot of my work is starting to get processor-heavy. It's just a pain in the ass waiting for the darn computer to compute a couple particle simulations.