Neural Networks That Draw: What They Are and How an Artist Can Use Them in His Work

A neural network is a concept from biology. A neural network is very similar to the brain: there are neurons and connections between them. Neurons receive, process and give out information, and connections transfer it. The difference between the human brain and a computer neural network is that neurons in the brain can independently form stable connections, when a person learns something new. Whereas neurons in a computer network are not able to learn and form connections on their own. They can only receive, process, and give out information on command.

For a neural network to work, it has to be taught. It’s similar to the way one explains new things to young children. You show your child pictures of animals and name them: cow, cat, dog, camel. Then the child shows you the correct picture when you ask him to “show me the camel”. This is also how a neural network works. There is a complicated process inside processing the information. But the principle of learning and subsequent command execution is the same.

The biggest misconception is that neural networks can learn on their own. Remember Ultron in the Avengers movie of the Marvel universe?

He gained access to information, went online and became the most powerful intelligence. And then he acquired a physical shell. Now, Ultron’s self-training is fiction. All modern neural networks are trained by humans. And the algorithm of their work is written by the same person.

Each neural network has its own task and its own algorithm. For example, Google picture search is also a neural network. You show it a picture and press the “find” button. For a neural network, this is a command “find me the same one, or anything like it”. Google has a preloaded database of pictures, and the network takes yours, runs it through its database, and compares it. And eventually gives you everything it thinks is similar. How to run through the database and what pictures are considered similar, the neural network in advance taught the person.

But the DALL-E network generates its image from the elements it knows from the text description. You write “astronaut riding a horse, pencil drawing”. The network itself searches for suitable images, combines them, adds the specified image style and outputs the finished picture.

Disco Diffusion

It’s a neural network from Google with an interface, which is as user-friendly as 22Bet Canada. It creates pictures based on text descriptions: you need to think of a theme or a story, and ask it to the network. In response, it draws a picture based on its knowledge of the given plot. If you repeat the same subject again, the picture will be different.

Select the number of steps in the steps column. This is how many pictures will be used for the final result. The more steps, the sharper the picture. But don’t choose over 500, otherwise the picture will be too detailed and take a long time to be generated. Usually 250-300 steps is enough. In the column width_height you can choose the size and format of the picture.

Next, in the prompts section enter your request – what you want the neural network to draw. Besides the theme itself, you can either enter the name of artist/artists, whose style the neural network is going to rely on, or you can enter the name of artist/artists on Artstation.

Then go back to the menu at the top of the page and go to: Run Environment -> Run All. Or press the key combination Ctrl+F9.

The network will ask you for disk access and get to work – it usually takes 1 to 3 hours per image.

Artbreeder

It mixes different images together. You can take several portraits and get a new face. And you can change the gender, age, mood, facial features. You can take portraits only from the database, you will not be able to load your own images.

To generate a new image, click on the green plus sign in the upper right corner of the screen. Then in the window that will open, select the theme of the future picture.

Then click on the “Add pattern” button and select the images to be processed. And then play with the settings to get an interesting image.

Nvidia GauGAN

This is actually not one neural network, but two. One of them creates images very similar to those in the neural network knowledge base. And the other one determines whether it is unique or not. So it’s as if the two neural networks are constantly competing with each other. That is why this neural network is called generative-adversarial.

Choose a picture subject and draw on the left side of the screen.

When the image is ready, check the boxes at the bottom of the workspace to agree to the terms of use. Then select an image style and click on the right arrow button. On the right side of the screen you will get the result.

EbSynth

A network for styling and changing images or videos. To do this, you need to take one frame, save it, redraw it the way you want, and run the command to process the video according to that frame.

The process is quite time-consuming and requires preparation. You need the video and storyboard in .PNG format. Select one frame, process it and save it to a folder. The program will synchronize the processed frame with the video.