In the technology world, a 100% true-to-life image is the Holy Grail. It’s a holy grail that cannot be achieved without artificial intelligence (AI). The problem: we can only achieve it with a significant amount of data.
The question is: how do you get hold of some of that data?
If you want to make an image that looks like anything else on the internet, you have to delve into the trade secrets of Facebook and Google. But if you want to produce images for your marketing materials, you can automate them entirely with AI. That’s what we’re going to talk about today.
How does it work? Let’s look at an example: imagine you want to make a “Horror Image.” This image should be scary, but not too scary as to draw attention away from the message; it should have enough depth and detail so that it appeals to people who are looking for something more abstract than an object on their screen, but not so much detail that it hurts the message.
You could go through all your imagery files and manually select all the parts of your image that are important in relation to its meaning or message, such as caption text or background photos (like those used in social media), but then again, if there’s only one photo or caption text file in your images folder, then you could use AI instead.
AI is able to produce images directly from text to create an entire image.
The first time I used the term “AI”, I thought it was a synonym for “artificial intelligence”. But my friend, who is a computer science professor and a photographer, pointed out that the word “artificial” itself can be taken to mean an intentionally made thing. A lot of the most interesting projects today involve artificial intelligence (AI) in one way or another.
A couple of years ago in a post called “AI-Assisted Image Editing with Adobe After Effects” by David Gilson at Creative Cow , he talked about how he used AI to automatically add some text to images he was dealing with. He found that this produced more interesting images than simply letting the images do what they were already doing ( adding text ).
The problem is that, if you want to use image editing tools because they are difficult to work with (or you are just not comfortable with Photoshop), it would be very hard to get them to do things like add text when you don’t want them to. There has really only been one successful project like this: the AI-assisted image editing tool called Pikup . Pikup is developed by Vagaas , who has built an open source software platform for AI-assisted image editing known as H2O . H2O is available for both Windows and MacOS since early 2014 and has quickly become popular as a tool for creating custom retouching workflows using multiple types of image editors (such as Photoshop).
I’ve been using H2O quite extensively over the last month and have had many mixed experiences with it. Some of the things I’ve done which work well are:
1) Adding text on existing images
For example, here’s a picture of my son which I found myself wanting to colorize. It was already colored in Photoshop, but I didn’t want Photoshop adding text on top of it so I found Pikup and added some text there:
Pikup showed me where I could find the line between “colorization” and “text”. By default, texture maps are added at different points along the lines within an image. For example, here’s how it looks without any texture map:
In this case I wanted something more like this… where there is no texture map at all yet H2O does its best to make sense out of what’s there (the lines). You can see from above that there are no lines between colors in
But there is a way to have more control over parts of the image – adding AI generated elements to existing images.
This is a simple example of how to use AI to create useful parts of an image, without the need for any deep neural network. It’s not just for photo editing either, as we can use it to create more complex shapes.
What you’ll do here: Create a basic shape from text by using the letters in the text as a starting point and then adding path-edges and other features. In this case, we want a triangle – which we can achieve by creating three points on each side and then using the 3D camera to draw them.
You can find out more about these techniques in my book, The Art of Images (it’s free!).
This can be used for branding, product placement, or any other purpose where you want variety in
When you have a great idea to market your product, it’s often possible to generate an image of it (in a similar way to how you can do video-generated content). The starting point is text and the result is a nicely designed image.
The problem with this approach is that if you want to use an AI in generating images, you should also use an AI for evaluating the image quality. Let’s say your keyword has “artistic” in it and the keyword “art” has been used very frequently. The words “beautiful”, on the other hand, could be more likely to appear in search results for “artistic art” than for “beautiful art”. And since the AI is not able to make a judgment about that, it cannot be used as a quality control tool either. This is where controlled with AI comes into play:
For example, let’s say there are two keywords with images that we want to generate: one generated by an AI and one generated by us (if we don’t use an AI). We only need to check the quality of each generated image ourselves – so no additional layers of code or content are required on our part – and if they pass our test then we can continue generating images with the same quality as before.
This can obviously be applied whenever you have some control over what’s inside of your product or service rather than just when you are optimizing for Google or iTunes rankings.