Many people say that this year is the “first year of AI painting”. First of all, Disco Diffusion is popular and everyone knows it. From Text-to-Image (using text to generate images) to develop the community and the creative design industry, it has become popular in the eyes of ordinary people.
People are keen to put two completely incompatible objects, such as the words “Da Vincic” and “iPhone”, into the AI program, and then wait for the picture to be drawn layer by layer.
It was a kind of “unpacking the lucky bag” experience. For people without any art foundation and painting ability, most of AI’s “melting stems” pictures are amazing enough. Even if the effect “rolls over”, they can continue to be optimized by adjusting the descriptors.
Then, the AI painting tool Midjourney also became popular. Different from Disco Diffusion’s simple interface full of English and code, Midjourney is directly on the Discord channel. The process of entering commands is no different from sending messages to others. What is even more surprising is that it usually takes 60 seconds to generate paintings. about.
Then, OpenAI’s DALL·E 2 came out halfway. Unlike the previous two, which are good at “conceptual style”, DALL·E 2 is more “realistic”, and can generate 10 pictures in less than 60 seconds. If you are not satisfied, you can also wipe off parts Regenerate… In just a few months, the title of “The Strongest AI Painter” has changed hands several times.
Google couldn’t sit still either. At the end of May, it published a paper to introduce its own contestant, Imagen, directly calling DALL·E 2, claiming that Imagen has “unprecedented realism and deep language understanding”, which is not yet open.
In the past two months, I have dealt with the first three “AI painters” frequently, testing descriptors, training robots almost every day, stepping on a lot of pits, and turning over a lot of cars. But at the same time, a lot of masterpieces have also been obtained.
This time, I will compare their characteristics, user-friendliness, etc., and organize their URLs, as well as some simple operation methods.
For ordinary users, they are powerful tools for figurative imagination; for professional people, if they are linked with other tools, they can have endless imagination space.
Disco Diffusion: The resulting graph is the most artistic
The process of creating paintings by Disco Diffusion can be roughly divided into these steps: open the program; set parameters such as the image size, the number of process maps, and the number of generated images; write the description words (Prompts) in English, and the format is roughly “painting type + object” (There can be more than one) + painting style setting + some rhetorical words that limit the role”; then start running, waiting for AI to calculate the painting.
Generally, you need to wait half an hour, if you stare at the screen, you will see the image from being full of noise to gradually becoming clear and detailed.
During use, Disco Diffusion may prompt you to free up enough running memory on the computer, but because it runs on computing resources such as GPU provided by Google for free, it does not require high hardware requirements for the user’s computer. Open the browser to execute can.
Disco Diffusion itself is a free open source software, but if you want faster drawing speed, you can buy a Google Colab membership to allocate faster cloud computing resources.
In addition to only entering text to let the AI play freely, you can also put an Initial Image in advance to constrain the AI’s creation.
For example, I first made a base map with tree outlines and green color blocks (left), and then operated, Disco Diffusion will play in this big frame, and the finished product is the right image
The graph generated by Disco Diffusion can theoretically be commercialized. Its program is based on the MIT open source agreement. All network users can use, copy, modify and even sell the graph for free. But I think there is still a risk. The risk is mainly due to the fact that your descriptors will lead to plagiarism disputes.
When you use an artist with a distinctive style (especially a living artist), and a commercial work as a keyword, please don’t use it directly for commercial use.
Midjourney: Less “cross-border”, more “obedient”
Midjourney is still by invitation only.
In order to test the effect of Midjourney, I copied the keywords that I “feed” to Disco Diffusion before – “starry sky”, “sunflower”, “Van Gogh” – and pasted into it.
When I saw the finished product, I had an intuitive feeling: Midjourney’s imagination is not as wild as Disco Diffusion. But if I consider it from the perspective of auxiliary creation, I would be more inclined to use Midjourney, a more “obedient tool”. After all, no creator is willing to give up creative control to AI.
The advantage of Midjourney is that it is fast. The software generates graphs very quickly, about 60 seconds for one image. If you are not satisfied with the finished product, you can also enhance the details almost instantly, or extend the changes.
Midjourney hooked up to the messaging software Discord. After entering “/image” in the dialog box, enter the descriptor in English and press Enter. This process is like chatting with AI.
After 60 seconds, you will receive 4 pictures in the dialog box. If you are not satisfied with “Figure 1”, you can click the “U1” button to add details, and press the “V1” button to extend the changes until you are satisfied.
So, I took Midjourney to produce “McDonald’s in the 19th century” and “Migrant workers in the 18th century”:
The reason why Midjourney is a “productized” Disco Diffusion is that its interface is more friendly, and the other is that it also has a built-in creative community, you can see what descriptors players use to produce what kind of painting. This is a very valuable “painting style” database, which is very suitable for “copying homework”.
For example, I tried to generate the scene of the episode “Bad Journey” in “Love, Death and Robots”, referring to the descriptions of the two artists in the picture above, and then I produced a satisfactory painting:
“Copying homework” further lowers the threshold for producing decent works, but on the other hand, it also loses a lot of the fun of exploration. Don’t let game tips ruin a good game.
In terms of copyright, if you are a free user, the copyright of the image belongs to AI, and after paying $30 per month, you can use the image for commercial use. But at the same time, if you make a profit of more than $20,000, you need to give Midjourney a 20% profit.
DALL·E 2: I cut Van Gogh’s hair, I turned the elephant
I went back to being a hairdresser and styled Van Gogh with the DALL·E 2.
I waited more than a month before I got the qualification for the internal test of DALL·E 2. If Disco Diffusion is better at depicting atmospheres, landscapes or concept art, then DALL·E 2 is better at realism.
“Can the elephant turn around?” I took this super unreasonable demand as an example to try the realistic ability of DALL·E 2.
It turned around.
Also let the elephant do other things. For example, let an elephant swim in an aquarium:
Let the elephants dance with the sharks:
Let the elephant go wild on a Harley:
Let the elephant be weighed by Cao Chong:
This result leaves people speechless.
It is no exaggeration to say that this is the best AI drawing tool I have ever used. The operation is simple enough, the degree of completion is high, and the speed is fast enough to be a search engine: 10 images (1024 × 1024) are generated in less than a minute, Changes can be extended infinitely, and can even be partially regenerated by erasing. You can keep helping Van Gogh “cut his hair”.
In terms of copyright, OpenAI, the organization behind DALL·E 2, has listed several strict restrictions: the copyright of pictures is ultimately owned by OpenAI; it is only for personal learning and exploration, not for commercial use, and cannot be used to make NFTs; it cannot be published on social media Too realistic faces produce results, and there is a risk of portrait infringement.
OpenAI also claims to have banned AI from remembering the faces of celebrities, as well as circumventing racial and gender stereotypes, among other things.
Before waiting for the qualification of DALL·E 2 internal test, I found a “stand-in” – DALL·E mini, which is a demo made with the first generation DALL·E. The production speed is fast, but the picture completion is not as good as that of DALL·E E 2.
Generating an image is just the first step
“Can you make them move?” I looked at the paintings sent back by the AI and began to think of a way:
The completion of AI-generated images does not mean the end of creativity. If you take it as one of the links, and then connect to other creative processes, the imagination space is huge.
I’ll show the illustrator again Nerko Creativity: He uses Midjourney to generate the material he wants, and then assembles the parts.
Another example of how its done.
I generate with the AI (midjourney) many parts and shapes that may be interesting, I cut them and prepare them to later make the composition in photoshop. Some overpaint, shadows, lights and that’s it. ???? pic.twitter.com/0EnHRLJRjq– Nekro (@NekroXIII) June 12, 2022
In his hands, AI is a kind of “productivity”. Selection and synthesis are still under his full control. He had been illustrator for 15 years before using Midjourney.