How to use a neural network to generate Midjourney drawings: get the request right
How to use a neural network to generate Midjourney drawings: correctly compose a query
Midjourney is the main neural network of the summer and a reason to once again see the progress of technology.
Dall-E 2 is still closed to most users, and Dall-E Mini is more of an entertainment service than a serious tool. Against this backdrop, Midjourney stands out for its affordability and impressive results. The pictures that it generates from a text description can be confused with “human” art, and more and more people are starting to use the neural network.
There are a huge number of parameters in Midjourney, and many of them can only be reached by experience. Here a problem arises: the number of free attempts is limited to 25. Therefore, we tell you how to correctly formulate a request to a neural network and not exhaust the limit for not very high-quality results.
- How to try the neural network yourself
- How to correctly formulate a neural network query
- What are the services for compiling the perfect request
- Final memo
35% discount on all courses of the Textbook
How to try the neural network yourself
How to access Midjourney. Right now Midjourney can only be used as a bot on Discord. So you will have to register on the platform and join the official server of the neural network team. Now there are already almost a million users, access is open to everyone by invitation link. But recently, users who have not been active for a long time have been removed from the community: you may have to re-join over time.
In August, the developers allowed the Midjourney bot to be added to third-party servers. The main thing is that they have up to 10 thousand subscribers. So you can add the bot to your own community or to another server where you are among the administrators.
How the Midjourney discord server works. There are several main channels: announcements from developers are placed in Announcements, technical problems are reported in Status. At the same time, I advise you to familiarize yourself with the rules in the Rules channel so that there are no questions if you are suddenly banned for something .
But at the heart of the server are spaces for generating images. There are Newcomer Rooms channels on Discord, inside of which there are hundreds of Newbies chats. It is in them that users create pictures. Choose any of the chats, ideally - so that not too many people are sitting there at the same time. You will see all their attempts to create a picture using a neural network, and they will see yours.
During active hours, hundreds of messages per hour can appear in the channel, and your work will obviously be lost. The "Mail" button in the upper right corner of "Discord" will help here: the bot mentions you every time a picture is generated, all notifications from it can be viewed in a separate window.
How to generate your first image. Log in to one of the Newbies chats and enter the /imagine command in the send message field. A Prompt box will appear: anything you write after it will become part of the query.The description of the picture for the neural network is better formulated in English. Other languages Midjourney understands much worse.
When you formulate the request, click the "Submit" button. The bot will start generating an image. You will literally see live how the vague outlines turn into four pictures. Usually, the generation takes no longer than a minute, but it all depends on the load on the Discord server and the details of the specific request. By the way, some parameters are set by the bot initially - for example, the generation speed. You can view or change the settings by sending the /settings command.
The four resulting pictures are not the final result. The buttons "U1, U2, U3, U4" and "V1, V2, V3, V4" will appear below the collage. The first parameter is upscaling. Choose the most successful picture from the four, and the neural network will improve its resolution, and at the same time add details. The second parameter allows you to create four more variations of a particular picture from a collage - in case you like the result, but want to see what else the neural network will do in the same vein.
How to save an image. First tip: don't save a picture or collage directly from a chat. Click on it and select "Open Original". So the image resolution will be higher, which one depends on the parameters you have already set. It's best to save the whole collage if you want to share all the options, or already a picture with the maximum level of upscaling.
The second tip: you can ask the bot to send you pictures in private messages. To do this, respond to a post with the result from the emoji bot in the form of a letter. It can be easily found in a search for :envelope:.
And your results are stored on the Midjourney website - you need to log in through Discord.
What are the restrictions. There is a limit for free image generation - 25 pieces per user.It is important that this number includes not only original requests, but also requests to improve the detail or make variations of the already received picture. A Discord bot will alert you when your attempts start to come to an end.
Midjourney has a paid subscription: $10 (
Of the other limitations: there are stop words with which the generation does not work. The community collects them in a separate place.
How to correctly formulate a neural network query
In favor of neural networks, it is often said that it is enough to choose the right query and click the “Make a masterpiece” button. This is only partly true: technically everything is correct, but it can take hours to find the perfect query. This is especially true in the case of Midjourney - the limit of 25 free generations greatly increases the price of each mistake. You do not have time to look back, as the limit is exhausted.
Therefore, the developers of the neural network and the community on Discord and on Reddit are engaged in the fact that they make rules and advice when making a request. Some of them were formulated by the Midjourney team on their official website, users came to other life hacks by experience. Here are some of them.
Remember that you are talking to a neural network. This should be reflected in the request. Do not write "Favorite place in Moscow in 2100", but be more specific, for example, "Red Square in Moscow in 2100". If you can refine the wording and add descriptiveness, do it. The developers give the example of the query "Monkeys are doing business" - it is much easier to write "Monkeys are sitting in the office in business suits."
Also avoid the words "not", "without", "excluding", "except" and the like. The neural network will get confused in the query “The man is not in the blue hat”, but the “Man in the red hat” will be generated without problems.
Specify the parameters correctly. The community has come to certain rules on how to make a request to the neural network. Do not try to formulate a complex sentence with an artistic description of the picture. It is much more efficient to separate parts of the query with commas or two colons "::". Highlight the main objects, style, parameters and short commands for the algorithm in this way.
Useful short commands to consider
--iw. Midjourney allows you to upload pictures as a reference. You can specify its "weight" - the more, the more it affects the result. The default value is 0.25.
--ar. A shortcut for the aspect ratio parameter, or "aspect ratio". By default, pictures are generated in a ratio of 1:1, but this can be changed - for example, to 3:2 or 16:9.
--s. styling level. A rather abstract parameter that can be set in the range from 0 to 60,000.
--q. Quality level. By default, it is set to 1, you can set it to 2, then the picture will have more details. Previously, there was a maximum parameter of 5, but it was removed due to the high load.
--chaos. The level of "chaos" in image processing can be increased up to 100.
--wallpaper. This is how the neural network understands that you need a detailed and high-quality image. This command is more convenient and efficient than the "4k", "ultra detailed" and other options.
--v. Allows you to specify which version of Midjourney will generate an image on request. The default is the latest version, the third. I didn't see much point in changing it.
--video. Saves a video of the image generation process. Add to the request, and then ask the bot to send the video to private messages using emoji letters.
The Prague Media School recommends the following sequence in the request: "Object description :: main style :: additional parameters :: --short commands". In my Corgi example, it usually went something like this: "Corgi biker on red motorcycle :: in soviet style, mosaic :: --q2".
If you are just getting acquainted with the neural network, it is better to start with the usual description of the object, and then gradually complicate the formulations and look for successful combinations.
Set the style. Midjourney has been trained in a wide range of art styles, the work of specific people, different aesthetics and time periods. Much depends on the imagination of users. Don't be afraid to add details: colors, lighting, materials. In my case, the Corgi turned into a photorealistic portrait, then into a 3D render , then into origami, then into a tattoo, then into part of a Soviet poster.
There are thousands of stylistic details that affect the final image. Some of them are described on Github: there is even a more or less illustrative example with a sphere that changes depending on the parameters. If this is not enough, the community collects various visual styles with examples in a table.
If you're into photography, try putting your skills to the test when making a request. You can set the type of camera, lens, object distance, focal length, shutter speed. By the way, in an attempt to imitate photographs, users often write the parameter Photorealism. On Github, it is advised to write Photograph or 35mm photograph instead . Another option is to write the name of your favorite famous photographer.
In general, regarding styles, the Midjourney team advises the following language:
- in the style of - if you want to add some style or aesthetics;
- by x, where x is the artist whose work will inspire the neural network.
Do not be afraid to mix different styles, the main thing is not to go too far and not to enter mutually exclusive requests. A neural network can generate a hand-drawn cyberpunk corgi, but there are problems with both neon and black-and-white corgi.
Let go of the situation periodically. What I noticed while using Midjourney: if at first users got by with short formulations, now multi-level queries for several lines prevail. Yes, this way it comes out much closer to what you imagined in your head, but the element of neurocreativity is lost.
More than once or twice, I came to the conclusion that in the end I like a simple request of three or five words much more than a description of a dozen parameters.
What are the services for compiling the perfect request
It is difficult to remember all the filters, parameters and recommendations. Therefore, enthusiasts began to create "prompters" - small services that help formulate a request. They work in much the same way: the user himself makes a basic description of what the neural network should draw, and selects additional parameters from a ready-made list.
Important: the service will only give you a ready request, which then needs to be passed to the bot. On sites-prompters, you cannot immediately generate an image.
Prompter. A service developed by third-party developer Shane McGuian. Unlike other resources, this is literally a google spreadsheet open to everyone , which is constantly being improved and updated. Almost every option has a short explanation, and there are a huge number of styles to choose from. But, unlike similar services, the settings lack visibility.
Promptomania. At first glance, the most advanced of the prompters. A huge number of parameters: they are conveniently categorized, almost every setting is illustrated.
But in such a detailed query editor, it is easy to get confused, and a huge amount of detail leads to slurred results. But the tool will become a powerful assistant if you understand it.
Phraser. The most convenient service. It does not have a lot of settings, but they can be selected step by step. Because of this, the eyes do not scatter, but over time a small number of parameters affect.
To use the service, you must first register on the site.
Noonshot. Pretty detailed query editor.
In addition to style, camera and color settings, you can set the "intensity" of the parameter. This is the most difficult thing: the slider changes the final image in a rather non-obvious way.
Final memo
- Midjourney is available for free to Discord users. It is best to test the bot in the official neural network community - you will be able to see what others are doing at the same time. But the bot can be added to your server.
- Remember about the limit of 25 free generations. You can take a paid subscription, but Russian maps are not supported. Alternative: create a new Discord profile, but if the community administration finds out, you will be banned.
- Generate pictures in one of the Newbies chats. Enter the /imagine command and formulate a query.
- Decide on the object - what or who will become the main character of the image. Immediately write down the details that you need: color, camera settings, materials. Add style, it's best to do it through the following commands: in style of or artwork by.
- Remember the short commands, with their help you can set the image detail, aspect ratio and other parameters.
- You can make a good request with the help of services-prompters. They clearly demonstrate what opportunities the neural network hides.
We post circles, beautiful cards and news about technology and pop culture in our telegram channel. Subscribe, it's great there: @t_technocult.
Коментарі
Дописати коментар
Олег Мічман в X: «Donations and support for media resources, bloggers, projects, and individuals. https://t.co/HPKsNRd4Uo https://t.co/R6NXVPK62M» / X
https://twitter.com/olukawy/status/1703876551505309973