Further Experiments with Deep Dreaming towards #HANDDRAWNPHOTOs

Yesterday I did a series of panoramas with two different cameras on the leap track in the southern part of Jells Park. This was primarily to test out the panorama stitching in the new version On1 Photo RAW 2019.2 (trial version). I was quite happy with the results, which managed to capture the quintessential nature of the Australian bush. This inspired me try some further experiments in using Ai methods to refine my #HANDDRAWNPHOTO idea. Basically I’m investigating different styles of photography and post processing that might allow the hand of the artist, his mark making/style to shine through and become part of the photograph.

So the first step for me was to create a simple semi-abstract sketch of just the shapes and textures I could see between the trees in my panoramas. This was made with sharpie pens, in simple blue, purple and black hues. (i wasn’t going to try and use the colours from my sketch just the tones). This is very loose and not too accurate but that was what I wanted.

I used this sketch as my default style in the Deep Style module of Googles Deep Dream, This involves a bit of time to process because the software has to create a neural network which will understand the key style elements, lines and shapes and how I draw them. The original photo is uploaded image at the generate screen. I have been constantly tweaking the settings under the style tiles. The important setting in these examples is preserving colours (from the original photo). Its trial and error from there and can take time, so enjoy it, have a coffee.

This pair shows the original photo above and the resulting deep dream below.

I think this is starting to look what I am after.

PhotoProject :: My #HandDrawnPhoto approach to #AIart

The success of the Belemy portrait auction, has encouraged me to compare Obvious’s DAG approach to my use of Google Deep Dream algorithm. Deep dream uses a convolutional (foreword chaining) neural network to find and/or enhance patterns in an image. This can be applied in two ways Style & Inception (google’s terms). The inception process tries to find the characteristics of a given object/pattern within a photo and tends to produce surreal images. Style on the other hand  seeks to put characteristic patterns back into a photo based on features the trained neural network has recognized. This is the technique I have become very interested in as a way to put my characteristic marks (and sometimes colours) back into a photo. Hence the name Hand Drawn Photos for my approach. The method does have superficial similarities to Onvious’s DAG Approach.

Here I am using my sketch for inktober as the image to be processed and a photo of the same trees as my training image used to build the neural network (AI). The Photo here is acting like the discriminator in Obvious’s approach. The result (below) is tonally similar to the photo and kind of a messy squiggly version of something I might sketch.  But…

In this, my preferred approach, I am using the photo as the image to be modified and my sketch as the image to train the neural network and letting the system modify the photo to “find” my characteristic mark making (aka style) and modify the photo. In this case I am taking the colours from the photograph, but that is not essential.

I feel this second result is much cleaner and fresher, closer to something creative and artistic, Although I am strongly guiding the outcome by using my own sketches (as opposed to an automated algorithm). So the big question is “Do you consider this #HandDrawnPhoto to be legitimate AI art? I’m warming to answering yes.

If you are interested here is the link to my profile and public images on Google Deep Dream website. (Its free to join and use)

Love to hear your views,  please leave your comments here.