Generating Art With Artificial Intelligence Powered Applications

How good is the art generated by an AI? Find out now!

Playing chess, driving cars, cooking food. Only a few decades ago, one might have thought it to be impossible for a machine to perform such tasks. Creativity is considered to be something very unique to humans. Can AI ever truly be creative? Can AI make original art as humans can? Turns out that it can and it is pretty good at it.

In 2018 an art piece generated by an AI-based algorithm was sold for about half a million dollars. This is achieved through Generative Adversarial Networks (GANs) where two models are trained simultaneously; a generative model G that captures the data distribution, and a discriminative model D that estimates the probability that a sample came from the training data rather than G. The training procedure for G is to maximize the probability of D making a mistake. That’s as much as technical we will get with this article because that is not the focus point. We are here for some fun and to get amazed by what cutting edge AI of today can do.

𝒎𝒊𝒏 𝑮 𝒎𝒂𝒙 𝑫 𝔼𝒙 [𝒍𝒐𝒈 𝑫 (𝒙))] + 𝔼𝒛 [𝒍𝒐𝒈(𝟏 − 𝑫(𝑮(𝒛)))], Portrait of Edmond de Belamy, from La Famille de Belamy (2018). Courtesy of Christie’s Images Ltd. (An art generated by AI)

Let us create some art, shall we?

Researchers and developers have built quite a bunch of tools and applications that can enable anyone to easily create art using AI. We are going to explore some of those applications and experiment to see the potential of AI.

Most of these software work on the basic ground of GANs with variations of their own. With that being said, let’s get started!

ArtBreeder

This is an online application in which you can merge pictures and create a hybrid of the two. How it actually does this is based on this paper by lead author Andrew Brock. Here, we will explore the tool from an application point of view.

• The user has an option to choose the category of the picture they want to generate.
• Post that they can choose from the database and combine as many pictures they want. The results are created extremely fast.
• The image generated is based on the most prominent features of each input picture.
• The major drawback of this app is that one can not import their own work and is restricted by the database. This might lead to a mixed conclusion of the algorithm depending on memory rather than AI.

Here is some stuff we tried to generate using this tool.

The bottom image is generated by mixing the top3 images. You have options to adjust the style and content of the image.

Here, we created a new anime character by the fusion of the above two anime characters

Looks intriguing? Visit the below link to generate your own combination of images!

Artbreeder

Deep Dream Generator

The Home Page

Ever wanted to convert your images into a dreamy sequence of pictures? This might be the tool just for you.

• Deep dream generator utilizes the base model of deep dream for its application. The main algorithm is based on the groundbreaking inceptionet paper drafted by Google, the code of which is open-sourced.
• Instead of fusing two images, a different style of the image is generated with a variation in the texture.
• You just need to upload your own image of choice, select the option for what style (deep style, thin style, or deep dream) and texture (there are quite a bunch of choices to choose from) you want to use, and then generate the new images.
• The deep dream is a direct implementation of Google’s Deep Dream which uses a convolutional neural network to find and enhance patterns in images via algorithmic pareidolia, thus creating a dream-like hallucinogenic appearance in the deliberately over-processed images.

Here is some cool stuff we have created using this tool.

The original input stock image

Generated image with deep style with laser fusion texture

Generated image with deep style and cold day texture

Generated image with thin style and texture 18

Generated image with Deep Dream style, with default settings.

Feels like some weird animals embedded into the image? Generating an image does take a couple of minutes, and there are many other settings to tinker around with. Want to give it a shot? Find out what you can generate with your images, here-

Deep Dream Generator

Deepart.io

This website specializes in fusing two images together. It works on the principle of very deep convolutional networks which is an improvement of the architecture developed by Andrew Brock. The image fusion generated is of very high quality. The website has a long waiting period before your results are generated, due to its image queue. As we are allowed to choose our own pictures which set a definitive that the algorithm is doing the actual work. One even has the option of buying art on this website.

We have used the same stock photo as before to generate a new image. The results are here (and it did take quite some time)-

Generated using deepart.io

Check this tool out here-

deepart.io - become a digital artist

Playform

This application provides a variety of options but we will be focusing mainly on their Sketch to Image feature. The software is run on GANs and high-end GPUs as claimed by their Github page.

Here’s how this works-

• You choose a texture style and then make a rough sketch of what you want to draw in their interface.
• On clicking the generate button, their program creates a bunch of images, using your sketch as the input, in the chosen style.

We have done some simple sketches, and the results are pretty impressive.

The input sketch we have made

Generated outputs by the AI using the Baroque style (Ain’t it impressive?)

A few more outputs in different styles

These are indeed some intriguing artworks created by the AI. The imagery is profound and feels to have inherent interpretability. Try it out yourself here-

Playform

GauGan

This is one of the state of the art facilities in development by Nvidia. In this app, every object is color-coded which makes it easy to design what you require.

• You can basically choose the objects and place them in the required positions of your image, and the AI will take care of the rest. You can also do a freehand drawing if you want.
• One issue is that there is a limited number of choices.
• Another problem one can notice is that it does not do well with the overlapping of objects.
• The GauGAN’s generative neural network model was trained on millions of images to be able to synthesize photorealistic landscapes, given a segmentation map.
• It learned a multi-modal distribution that enables it to produce different results for the same user input (or sketch). The method used is called spatially-adaptive normalization for semantic image synthesis is explained in this paper.

Among the existing technologies, it does perform pretty well. Let’s look at some of our results-

GuaGan Input (our drawing) vs Output. Notice that we have added several objects. Each color corresponds to one type of object/feature

Another input (created by us) vs AI-generated output (Ain’t it pretty cool?)

One key observation of the GuaGan is its photorealistic output, unlike the artistic ones we have seen before. Try it out yourself here.

Sketch — RNN Demos

The RNN drawing flowers (GIF Source- Google)

This sketch app basically completes the drawing that you have started. It is run by Google and is powered by a Recurrent Neural Network. This suggests that the algorithm works on a memory basis rather than a generative procedure.

• The user has an option to choose one object/model they want to draw
• And based on whatever rough sketches or lines you draw in the tool, the AI will try to complete them to create the model (It is pretty intuitive!).
• There is also an “everything” model, which tries to draw everything it can from the given sketch.
• One thing seen was that the algorithm rarely detects completion in the drawing and goes on to add unnecessary lines and curves.
• Also, animals such as cats are only restricted to the face and not the entire body structure.
• Lastly, the sketch lines have to be done at once as the algorithm immediately starts to work the moment the virtual pen is lifted.

Here are some sketches we tried-

Input vs output. The sketch-RNN completed the cat diagram

Input (just one vertical line) vs output for the “everything model”. The AI tries to create different sketches. One of the sketches is shown here

Seeing the AI in action in real-time is in itself quite impressive. Try out this fun tool here-

sketch-rnn

Creative Use Cases

From all the above examples, one thing is clear. AI has come a long long way in creating art. And there are so many different techniques available that can generate them almost equivalent to human talents, if not better. Here, we speculate on some ideas on using such tools.

• AI made paintings are already minting money across the globe. They are intriguing, difficult to explain, and often leave the viewer in awe. No wonder, the market for them will only increase.
• AI can mass-produce art. Something that takes an artist several days to create, AI does it in minutes. We can only imagine how it will transform this industry.
• AI can be highly assistive. We are no experts with art, yet we were able to create some interesting pieces. AI will be able to help anyone create good art, powered by their imagination and real-life images.
• With better and accurate models, there is a good possibility that AI will become powerful enough to create new character designs for video games, movies, etc. The possibilities surely go beyond that.

The Debate And Conclusion

The major technique of importing the various textures from one image to another, or its “style” is called style transfer. This technique is what actually creates the pieces. Art made by this process is now sold on many occasions and has a market of its own. The website noartist is completely dedicated to selling such art.

We are right now at the tipping point of an industry dominated by Human artists. Surely, AI will find its way, but if it will actually replace an artist is a question to ponder. Like with all other applications, AI is assistive in nature. We can expect it to evolve into a robust tool that will help artists work better and quicker. That might be the case for the foreseeable future.

But who knows? We may be just a breakthrough away before AI overtakes the art industry.

Editorial Note-

This article was conceptualized and co-written by Aditya Vivek Thota and Soumya Kundu of The Research Nest.

While we have not covered much of the technical aspects, we have hyperlinked relevant research papers, GitHub repositories, and resources for those who would like to dive deep into these topics.

Stay tuned for more diverse research trends and insights from across the world in science and technology, with a prime focus on artificial intelligence!