Let's start by enjoying a few pictures of the beautiful scene.

With their harmonious color palette and impactful compositions, most people might think they're the work of a sophisticated painter, but the images above are actually taken from YouTuber Quick-Eyed Sky's Disco Diffusion three-minute showcase video.

And Disco Diffusion is an AI authoring program that produces images when you type in keywords, it's currently an open source project for GitHub and anyone can play with it. But since it's not yet packaged like software, it's currently just in code form in Google's Colaboratory (aka Colab).

In recent years, artificial intelligence technology has developed rapidly, and in the field of artistic creation, such as AI writing, AI painting, and even AI arranging music has long been nothing new. With the promotion of more user-friendly human-computer interaction interface, now the general public can experience the charm of AI technology in a few easy steps by opening the webpage provided by developers.

This time we give you four AI creation tools that claim to be able to draw with their mouths (type in keywords and it comes out) to show you how to use them, and do a comparison on how to use them and how they come out, to see which AI is the easiest to use and has the best results.

Disco Diffusion: a hundred and one compliant B's

Designed by artist Somnai, Disco Diffusion has powerful keyword-to-image capabilities, and since the code is hosted on Colab, all rendering processes can be done online without relying on the user's computer to do the math.

▲ The creator's tweet

▲ Somnai's video made with Disco Diffusion

You may be intimidated by the amount of code when you first open the Disco Diffusion interface, but it is actually quite simple to use if you follow the steps below to debug it.

▲ A code-filled interface can easily turn people off

First, drag down the page to the third item "Settings", and on the right hand side "Basic Settings" we can set the quality of the generated image.

"The higher the value of "Step", the longer the computation time. However, artwork is to some extent a collision of chance and beauty, and more computations does not mean that the resulting work will look good.

The "width height" below refers to the image size of the resulting image, which can only be entered in multiples of 64 (e.g. 1280, 768), the larger the size the longer the rendering time.

There is a section called "Init Settings" in "Settings", and by entering the address of the image in "init_image", you can let the program perform operations with reference to a certain image.

Click the circled section below to upload the local image, then select the image in the list, find "copy path" in the right-click menu to copy the address, and then paste it to "init_image".

Next is the most important part of describing the content of the painting by keywords. Scroll down and find the "Prompts" item in "Settings" and see "text_prompts".

▲ Leave the square brackets and quotation marks unchanged in the statement

The default statement in the code in "text_prompts" is.

[“A beautiful painting of a singular lighthouse, shining its light across a tumultuous sea of blood by greg rutkowski and thomas kinkade, Trending on artstation.”, “yellow color scheme”]

Translated, this means.

"A strange lighthouse shining in a raging sea of blood, painted by Greg Rutkowski and Thomas kinkade, using Artstation as a reference library and a yellow color scheme."

• Beautiful painting" is a type of painting, but you can change it to "ink painting", "oil painting", "comic", etc.
• "a singular lighthouse, shining its light across a tumultuous sea of blood" is the content description that determines what will appear in your picture, so feel free to play with it here.
• The "greg rutkowski and thomas kinkade" is the style of the artist to which the picture is closer, you can enter more than one artist here.
• Trending on artstation" is a reference platform for paintings, i.e. the platform from which the AI mainly gets its reference, and the platforms that can be entered are "Facebook", "Pixiv", "Pixbay" and so on.
• Yellow color scheme" is the main color of the whole picture.

To change the drawing, simply adjust the appropriate keywords in the statement according to the format (note that commas, double quotes, etc. should be in English).

In addition to changing keywords, you can also set the weight of each keyword, for example when my painter's style is "Van Gogh and Monet (Van Gogh and Monet)", I can write "Van Gogh:3" and "Monet:2", which means the work is three points like Van Gogh and two points like Monet.

Sometimes we can also set the value to a negative number if we want an element to not appear in the screen.

▲ Please give me a painting of a cloudless sky that looks three parts Van Gogh and two parts Monet

Finally we can also set how many plots (variants) are generated at a time in the fourth item "Diffuse!" under "n_batches", i.e. if we enter 5 here, we end up with 5 plots (the more plots the longer it takes).

After setting the above important parameters, click the "Code Execution Program" button at the top of the page, and select "Run All" in the menu to start the program.

Free users are typically assigned Colaboratory Tesla K80 graphics computers, which are slower to render online, taking about an hour and a half to two hours to complete a graph of the default size and number of iterations.

If you can't wait, you can also choose to use your own computer's graphics card for rendering by "connecting to a local runtime", but you need to use a tool such as "Jupyter Notebook Server", so I won't expand on that here.

The next step is a long wait, and when you get to the bottom of "Diffuse!" you can see how your picture changes from a blur to clarity.

▲ The duration of the operation and the current step are displayed at the bottom of the program.

Based on the default statements provided by the previous program, I generated an image of a lighthouse at sea with a first-class image, which is perfectly fine to use as a desktop background.

Next I changed some more parameters and generated two more graphs with the following descriptive statements.

"A painting of a robot angel, flapping her wings in the night sky by Shirow Masamune, Trending on artstation, blue color scheme (A painting of a robot angel, flapping her wings in the night sky by Shirow Masamune, Trending on artstation, blue color scheme)"

▲ Default picture quality, 1.5 hours

"Cyberpunk station."

▲ 448*448 pixels, 45 minutes

First of all both of these images can be considered cut and dry. The "Mechanical Angel" image doesn't reflect the keyword "night sky", but the composition is distinctive and the angel shows a clear mechanical feel, so it's up to you to decide if the drawing has Shiro's authentic charm.

At the top of the 'station' diagram we can see elements such as trains and cables, and the red and green neon lights and screens are also typical cyberpunk elements, the beauty of which is that they are not very well finished.

NightCafe Creator: style migration wizard

Compared to Disco Diffusion, NightCafe is much easier to use.

The NightCafe webpage offers two smart drawing tools, the "Text to Image" on the left is a keyword-generating tool similar to Disco Diffusion, and the "Style Transfer" on the right can change a user's uploaded photo into a famous painting style.

Let's test the first feature first. Enter a keyword in the box below "Your text prompt", e.g. A city surrounded by machine, and select one of the 16 styles below.

Pull down to the bottom and select whether it is Texture First (Artist) or Composition First (Coherent) and click on "CREATE" at the bottom.

▲ The process of making a picture

If you are not satisfied with the default image size, you can also turn on "Show advanced options" in the keyword input screen. In this mode you can not only customize the size of the image, but also set multiple references for the base image.

The style transfer tool is also easy to use. Upload a photo, pick one of 64 famous painting styles, wait 1-2 minutes, and a stunning image is born.

Entering the previous keywords of lighthouse, station and angel into NightCafe Creator after removing the author and reference library, we get the following three images.

▲ Keywords: a strange beacon shining in a raging sea of blood, yellow color scheme.

▲ Keywords: a painting of a mechanical angel waving its wings in the night sky, blue color scheme

▲ Keywords: cyberpunk station

I think the most effective of all the images is the drawing of the angel. Although it is a more abstract route, the wings that look like the solar receiving wings of a satellite still reflect the combination of mechanical and angelic elements, and the movements of the character are also very vivid.

I think I'm not as interesting as the AI.

While the Cyberpunk station is moderately effective, the lighthouse figure is pretty generic, with a lot of similar lighthouses in the frame that can look like they were copied multiple times in an attempt to spread out the frame.

Cogview: Chinese keywords, eight images at a time

Cogview is the only AI tool in this test that supports Chinese input (word limit is 2-30 Chinese characters), developed by Tang Jie's team at Tsinghua University, with output time ranging from 1-5 minutes, and can generate eight images at a time, but the size is not adjustable.

From the fact that the web page will be translated into Chinese and then computed by typing in English keywords, it is inferred that the program should be trained in Chinese, which is rare among a bunch of similar AI that can only be typed in English.

▲ The team's paper published last year

In fact, this program is best at real photo generation, not artistic painting. For example, when we type in "the letter b made of soybeans," the AI gives us eight real photos that match the description.

▲ We have confirmed in various search engines that similar images are not readily available online

The previous NightCafe Creator only produced "artwork" like the image below, no matter how you switch the style.

When the keywords are painting-related, Cogview is a little less so, for example, when we want to generate a tower as previously described in the "Oil Painting Style" style, the eight finished images still look more like photographs than paintings.

The Cyberpunk Station has much better graphics, probably due to the fact that it's a building itself and the description itself lends itself to a realistic style.

I switched to an ink style when regenerating the lighthouse image to make it more like a painting. Cogview is the only one of the many AI's that actually reproduces the "sea of blood", but it's a bit weird.

The Mechanical Angels canvases come out all over the place, and all eight of the finished images are too abstract to evaluate.

But it is said that the handsome, humanoid images of angels with white wings that we are used to now were actually embellished during the Renaissance. The appearance of angels as described in the Old Testament would have been bizarre; after all, angels are beings beyond human cognitive ability, so how could they look like people?

▲ The seated angels depicted in the Old Testament

WOMBO Dream: beautiful interface, artwork heavy

WOMBO Dream This tool, which we previously covered in another article introducing AI drawing, has what I think is the best looking interface of the ones presented here, and the shortest build time of about 30 seconds.

The procedure is much like NightCafe Creation so I won't go into detail here, but let's get right to the pictures.

▲ Keywords: a painting of a mechanical angel waving its wings in the night sky, blue color scheme

▲ Keywords: a strange beacon shining in a raging sea of blood, yellow color scheme.

▲ Keywords: cyberpunk station

Since the program only supports up to 100 letters of the alphabet, some keywords we had to cut down. From the three images we can see that WOMBO Dream is very much a "drawing".

A "picture sense" is simply a work of art that one would think is drawn at first glance, unlike photography or 3D model posing. Of course, the program's own tarot card frame may be helpful in creating a "picture sense".

▲ Right-click Save As in the gallery to save the image without the frame

Summary

Now for a brief summary of the four AI drawing tools, first of all all the programs are web-based and free to use.

Disco Diffusion allows for the most complex keyword descriptions, with many self-setting parameters, and is a huge treasure trove to be explored, but takes a long time to map and has a relatively complex interface.

Both NightCafe Creator and WOMBO Dream can produce quick work, the former can set the size of the screen and can also accommodate more keywords, but the latter has a better looking interface and comes with its own picture frame.

The advantage of Cogview is that it supports Chinese keywords and has an advantage in this aspect of realistic photo generation, where it is not possible to customize the image size.

After a number of AI design tools became popular, some people started to worry about whether the work of painters would be replaced. After Disco Diffusion became a hit, the public Design Youth Lab asked a few design gurus to explore their feelings on AI painting, and one of the illustrators, Mad View, had this to say.

"If one day AI replaces more value, then I will try to seek out new value and the ship will come to a head."

The development of productivity brought about a change in the way of production. We all know that Impressionism was born from the invention of photography in the 19th century, and when people realized that no matter how they drew, they could never be more realistic than the camera, haziness, individuality, abstraction, etc. became precious.

▲ Monet's masterpiece "Impression of Sunrise

Now that AI painters have re-emerged to challenge our art world, what strategies will artists use to defend their territory? We'll have to wait and see.

• Disco Diffusion: https://colab.research.google.com/github/alembics/disco-diffusion/blob/main/Disco_Diffusion.ipynb
• Cogview：https://agc.platform.baai.ac.cn/CogView/index.html
• WOMBO Dream：https://www.wombo.art/
• NightCafe Creator：https://creator.nightcafe.studio/