Tablet menu

How to generate masterpieces in the Midjourney neural network? Step by step instructions


Midjourney is a neural network capable of generating impressive images according to specific requests. Today I share my experiments so that you don't waste your efforts.


The neural network is based on the Discord application. You need to have an account and get an invitation:

1. Go to the Midjourney website and click Join the Beta.

2. If you are authorized, then "Accept invitation". If you don't have an account, enter your details and click Continue.

3. After joining the Midjourney server, go to any of the newbies chats.


Each user is given 25 GPU-minutes for free. These are 25 pictures of standard quality, including revision. Signature: $10 per month = 200 attempts, $30 = unlimited attempts. *Cards of Russian banks will not work. I generate On the channel, you will see requests from other users and their results. To start, type /imagine and click on the suggested prompt command.


Enter a description for the image. In English, the results are better. Separate words with commas.


These cannot be used. In addition to the obvious ones, for some reason Tik Tok, Sex, and also Pin-Up are included. For the test, I took the request: "pepe the frog with mustache detailed picture". As a result, four different images appeared. They can be corrected - correct the description and generate again.



Below the pictures are additional commands.

The number indicates which image the action will be applied to:

1 - top left,

2 - top right,

3 - bottom left,

4 - bottom right.

U - enlarges and enhances the selected image;

V - generates four more similar images based on the original one;

icon with two arrows - options are regenerated.

If you decide to continue working with the image, press "U".

The system will display a larger, more detailed image.

You can also work with him, choosing from several options:

Make Variations - creates similar images;

Light Upscale Redo - easy redrawing of the original image;

Beta Upscale Redo - experimental upscale, initially generates higher resolution;

Detailed Upscale Redo - more serious redrawing of the image;

Remaster - completely redraws the image based on the source;

Upscale to Max - improvement to maximum quality;

Web - opens the image in full size in a separate browser tab. Calculating how further options will look like is not easy and the processed pictures are not always better.


Use the "Mail" button in the upper right corner. Go to the "Mentions" tab and you will see all your generations.

The bot can be asked to send the results in private messages. To do this, you need to respond to a post with emoji in the form of a letter. Go to the message and click "Add Reaction", search for envelope and select the appropriate icon.

To save the file, click on the image in the chat and then click on "Open Original". If you need one picture out of four and in maximum quality, then make an upscale (U button).

The results of the generation are also on the Midjourney website - the images will be in your personal account.


Complexity. Express yourself simply, otherwise the results may be far from desired.

Use popular images that are easy to find in pictures on the web. Exact query = clear result. If you do not directly register any parameter, it is generated randomly.

Change the successful requests of other users at your discretion, it is allowed to use successful wording.

Image style: drawing, computer graphics, realistic photography. You can also give references in the form of famous artists. For example, you can attribute to the request: by Claude Monet.


If the first images turned out to be unsuccessful, then it is better not to waste attempts to generate them again and start from scratch.


1. The first part describes what exactly should be shown.

2. The second part gives an idea of ​​the style.

3. The third clarifies the technical parameters. Parts must be separated by commas.


For example: pepe the frog with mustache detailled picture, christmas lights, unreal engine, cinematic lighting,--wallpaper. Commands preceded by two hyphens (in the example -wallpaper) specify technical parameters. They should always be listed at the end.


-v (example: -v 4) - algorithm version, values ​​from 1 to 4 are available (the first version is more suitable for abstract images and textures);

--beta - experimental algorithm that produces 2 results instead of 4 and takes longer to process;

-hd - an alternative algorithm that is more suitable for abstract images and landscapes;

--stop 10 - stops image generation at the specified percentage from 10 to 100;

-uplight - uses a more "sparing" upscale, that is, closer to the original image, similar to the Light Upscale button;

--upbeta --creates a higher resolution image initially;

--seed (example: --seed 1248) --seed value that is used to get similar results on each new request. You can enter any integer between 0 and 4294967295. If you don't choose a value, it is set randomly. You can find out if you send the result to yourself in private messages. --sameseed - the network will make all four images look similar;

-q (example: -q .25)

- image quality, you can specify a number from 0.25 to 5. The default is 1 (equal to 1 GPU). The higher the number, the more resources are consumed (remember the limit of 25 GPU-minutes);

-s - how strong the styling of the image will be. The default value is 2500. You can specify between 625 and 60000.

--video - saves a video with the process of generating an image that will be sent to private messages. You need to respond to a message with an emoji in the form of an envelope.

-iw (example: -iw .25) - allows you to set the weight of the reference image. Upload the image to the chat, copy the URL and paste it into the prompt after /imagine. The default value is .25, a value of 1 makes the weight of the image as important as the descriptive text;

-ar - sets the aspect ratio, for example -ar 16:9;

-no - allows you to exclude objects, for example -no cats; -wallpaper - a detailed and high-quality image is created, can be used instead of keywords like hyper detailed, micro details, 4K and the like.

The request can be formed much more simply, without the use of separators and with a minimal description. And in this case, the neural network can produce a decent result.

How to join two images

In the settings you need to set the Remix Mode, and then upload two images to Discord and paste the links in the prompt line. As a result, the neural network will make a mix of two pictures. The results may be unexpected.

Check out my top searches 


Last modified on
back to top