Midjourney

From NoskeWiki
Jump to navigation Jump to search

About

"Silhouettes of dancers" (v4)


NOTE: This page is a daughter page of: Artificial Intelligence


Asking for a herb themed website (v4)
"Avocado chair" (v3)

Midjourney is a program that creates images from textual descriptions, similar to OpenAI's DALL-E and Stable Diffusion. The independent research lab that produces Midjourney shares the same name. The tool is entered open beta on July 12, 2022 and became profitable by August. Currently, users create artwork with Midjourney using Discord (the chat platform) bot commands in the form:

/imagine A phoenix drinking wine, and with a red and black Chinese banner behind her. --ar 3:2

To use the platform you get a certain number (~20) free generations. To try it go to:


Note: Most of the screenshots here are from Midjourney version 4 (~Feb 2023) - version 5 was released mid March 2023 and is already a big step up in photorealism!


Videos


Midjourney Functionality

Midjourney is currently only accessible through a Discord bot on their official Discord, by direct messaging the bot, or by inviting the bot to a third-party server. To generate images, users use the /imagine command and type in a prompt; the bot then returns an image. Midjourney is also working on a web UI.


Pro Tip: Imagine Parameters and Phrases to Change the Style

In addition to the text description there are about a dozen imagine parameters you can enter at the end of the text with the typical `--` syntax. For instance --ar 3:2 to generate a 3:2 aspect ratio. It also pays to try certain phrases that will change the style of your image, such as "in the style of ___", "website, ui", "pixelart", "photorealistic", "3d render", "cartoon", "__ lighting", "portrait", "line art". The more examples you look at the more amazing things you can try.


Pro Tip: Getting Reproducibility

For reproducibility (to produce the same result using seed and version (eg: "--seed 1 --v 4"). For instance:

  • "a dancer exploding into the sky" --ar 1:1 --seed 1 --v 4


Pro Tip: Getting Max Resolution

Let's try that again:

  • "a dancer exploding into the sky" --ar 1:1 --seed 1 --v 4
    • Tip: If in newbie mode you'll lose it quickly, so open the inbox and click "Jump"
    • Tip: Try add argument --hd or --quality 4

And to get the max resolution:

  1. Now if you like the first panel click the "U1" button. This will upscale image #1.
  2. Click the result and "Open in browser".
  3. Save this image and it should be 1024x1024.... (~1.5GB)
    • If you want it bigger... you can try "Light/Beta Upscale Redo"... these will produce 1536 × 1536, but these usually lose quality, so it's not worth it.
  4. To make it bigger than 1024x1024 try letsenhance.io.


First part (v4)
Second part (v4)


Pro Tip: Making Consistent Characters

There are a couple of good videos on making a consistent character in AI.... they both involve:

  • Typing a really long prompt for a character you want including a name, "multiple poses and expressions", "100% white background", "--no outline", "multiple " etc.
  • Upscale an image you like... save the output... copy and paste each pose into Midjourney and get the URLS.
  • Click "..." > "Add Reaction" > "envelope"... and it will show the seed.
  • Make a prompt like: "ImageA ImageB ImageB, long-text-string --seed XXXXXX", but change the "multiple poses and expressions" to the pose you like.


Using Your Own Photo

If you want to use your own image and render your face as something different, you can use Discord, but it is easier to try uploading an image to postimg.cc. Then get the URL to that link and type:

/imagine https://i.postimg.cc/m2kJK4gk/Screenshot-2023-01-24-at-12-32-17-PM.png as a medieval warrior or frog.

Apparently, there are also ways to combine two photos. Using just one photo I feel it usually fails to render something that looks like you, versus Lensa AI which asks for 10-20 photos of you and is specifically trained to render your face in a set number of styles. At least with Midjourney, you can ask for any arbitrary style - just don't expect it to look like you as of version 3. Maybe version 4 will fix this.


Midjourney Use and Legalities

Midjorney has already been used to generate magazine covers and is said to be popular for quick design prototyping. David Holz, the founder and leader of the Midjourney team, says he sees artists as customers, not competitors of Midjourney. Since Midjourney training set may include copyrighted artists' works artists may request takedowns and for certain works to be removed from the set.


Some Examples

Here are some fun examples... all these examples were made January 2022, so version 3 of Midjourney. Expect it to get better and better with the loss cases.


Your Face Rendered Differently

Ever wonder what you might look like if you were a frog or princess? I do all the time. Midjourney will let you do that, but it's a little bit of hit or miss if the result looks anything like you at this point (version 3). Give it another iteration.

Midjourney results for: LINK TO PERSON cute cartoon .... and yes, an impressive feature if that you can link to a photo of yourself and ask to be rendered in any style (v4)
Midjourney results for: https://i.postimg.cc/m2kJK4gk/Screenshot-2023-01-24-at-12-32-17-PM.png as a medieval warrior .... doesn't quite look like me, but interesting! (v4)


Website

Wow, so if you use the right keywords Midjourney is amazing at prototyping ideas for a website. Pity you still have to write the HTML to try to replicate it.

Midjourney results for: Beautiful website for volleyball ... #EA772C, #060606, website --v 4 --stylize --q 2 ... notice the use of color codes and various flags - this is a professional at work, but possibly english as a second language because they have spelling mistakes and Midjorney easily corrects those. (v4)
Midjourney results for: A beautiful website for a website about consent to kiss - my more amateur attempt (v4)
Midjourney results for: A resume website ... (v4)


Specific Style Request

This is brilliant, if you have style of painting or art you like, ask for that style.

Midjourney results for: Mickey Mouse styled like Alex grey (v4)


Abstract Concepts

The most fun thing to try is often taking two concepts... like your friend's favorite animal and musical instrument.... or two crazy concepts like a "teddy bear" and "underwater spearfishing" and seeing what it produces.

Midjourney iterations for: Avocado toast as shown in the wikipedia article
Midjourney results for: A 3d render of a cartoon camel playing a harp (v4)
Midjourney results for: A camel playing a harp ... it went with photo realistic but camels can't play harps (v3)
Midjourney results for: Cat playing drums with city in background ... (v4)
Midjourney results for: Cat playing drums with city in background ... ... close up (v4)
Midjourney results for: Math owl. No doubt one of my favorites, and no - you won't find any of these pictures on the web - they are completely original words that take your input text and generate something new (v4)

Classic Art

Midjourney results for: in sea there was a cyclone, one ship, thunder (v4)
Midjourney results for: sea turn into building (v4)
Midjourney results for: 3d image of a jaguar .... ... well it didn't get the dancing part (v4)
Midjourney results for: is a photo realistic for a helicopter chasing a jeep ... I'm sure the person who requested this was surprised it rendered a helicopter rotor on the jeep, but this is the kind of oddness you can expect sometimes. (v4)
Midjourney results for: A cyber robot .... (v4)
Midjourney results for: Camera .... (v4)
MidJourney results someone else (not me) created for: Angel .... Pretty sexy image... interesting you can't put works like "sexy" into MidJourney, but that doesn't mean people are not finding the right words like "hourglass figure" to get seductive imagery back. This is a product of all art on the internet, and most of the pictures of women on the net are scandalous already. (v4)

Other

Midjourney results for: Photo of a man crying on a dance floor in a church surrounded by other people who are dancing and mostly dressed in white... I find it doesn't quite get photo realism here. Notice the extra fingers. (v4)
Midjourney results for: A phoenix drinking Pinot noir, and with a red and black Chinese banner behind her (v4)


AI's Impact on the Art Community

It's hard not to see these stunning images, generated by AI in mere seconds, wonder if the job of artists jobs is threatened. You could hang any of these stylized images on a wall and it will probably look just fantastic as an artist's painting - most people would never know. The big difference is that it can take an artists days to produce an original work, but with AI it takes mere seconds. But I personally don't think artists should panic just yet.

Here are some factors your maybe haven't considered yet:

  1. Art buyers value the original piece, not copies
  2. Artist usually take decades to build reputation
  3. People value time and narrative, over imitators
  4. AI art has flaws - especially with photo realism

I think a hybrid system of computers and humans is ahead and I've written about this in more detail at: The Impact of AI on the Art Community.


See Also

  • Adobe Firefly - Another amazing AI tool for generating and tweaking images.

Links