Midjourney ( https://www.midjourney.com/ , warning for moving background) is an image generation AI similar to DALL-E but with different ups and downs. Several people have remarked to me that my results are mysteriously better than theirs, so here is what I do and why.
First: Why use Midjourney specifically?
Pros:
- Cheaper than dall-e on a per-image basis
- Supports any aspect ratio
- Allows very high resolutions
- Exposes more fine-tuning knobs
- No watermark
- Easy to get extremely Aesthetic results
Cons:
- Very bad at hands
- Worse than dall-e in general at structural coherence
- Model seems to be more filtered/constrained than dall-e's
- Operated through a discord bot, which you'll either love or hate
- If you're on a non-corporate price tier, all your results are public
Aspect Ratio
Aside from the general prompt structuring advice that also applies to dall-e, the most important thing for good results in Midjourney is using the correct aspect ratio (image rectangle shape) for the kind of result you want. For example, a landscape should be in a landscape ratio (wider than it is tall) and a portrait should be in a portrait ratio (taller than it is wide). You control aspect ratio by ending your prompt with --aspect x:y (that's two dashes). The most important aspect ratios are:
--aspect 1:1square (the default)--aspect 16:9desktop wallpaper, HDTV (landscape)--aspect 4:3standard definition TV (landscape)--aspect 3:2photograph (landscape)--aspect 3:4portrait (sheet of paper, iPad)--aspect 2:3tall portrait (phone)
This is so important because Midjourney knows that some things are wide and some things are tall and if you don't give it enough space one way or too much the other, the results will be distorted and bizarre. Here's an example of correct use of aspect ratio leading to excellent results:
illustrated poster of a gender reveal tornado, tarot card --aspect 9:16
Tarot cards and tornadoes are both vertical, so I used a desktop wallpaper ratio rotated vertically (9:16). The result heavily emphasizes the verticality of the tornado, which has strong visual clarity.
And this is an example of using the wrong ratio:
cute watercolor painting of a shiba inu face --aspect 16:9
Dog faces are not very wide, and we didn't ask for anything else, so Midjourney has a panic attack and just draws... more dog. Asking for the exact same prompt in 1:1 gives much better results:
I enjoy that these good boys all look like they just broke into the jam jar. But if you didn't like that use of purple...
Negative Prompts
A common mistake that beginners make with prompt formulation is using the words "but", "no" and "not." Midjourney only knows what people say in captions and labels. Nobody has ever uploaded a photo and captioned it "a bowl with no apples in it." If you ask for this, you will get a bowl with a whole lot of apples.
Midjourney's solution to this is the negative prompt, which begins with --no. A negative prompt is a second, separate prompt that specifies what the image does not contain. I know this is confusing, but: do not use negative language in the negative prompt!
Correct:
--no apples
Incorrect:
--no there are not any apples in the bowl (this will be interpreted as no bowl either!)
You can chain --no and --aspect, the double dashes are used to distinguish them from literal uses of the word. So a bowl --no apples --aspect 16:9
When picking images, Midjourney will judge how well they do match the normal (positive) prompt, and how well they don't match the negative prompt. I have used the negative prompts --no eyes faces and --no writing to try to prevent unwanted weird visual artifacts; I had a particular problem with candy that had faces:
cute illustration of glitch candy --aspect 16:4
cute illustration of glitch candy --no faces --aspect 16:4
The composition of this specific result doesn't have a lot of candy but I think it's one of my most visually striking results in general. Sometimes Midjourney just surprises you.
Light Upscale Redo
When you upscale an image, Midjourney may go too hard and evolve the details into a bunch of weird gunk. If the preview thumbnail looked distinctly better, the solution is to click the "light upscale redo" button. This will take a more conservative approach and the result will be closer to a simple upscale of the preview rather than a "smart" one that adds more details. Here's an example where the details turned weird (she has a hand on her chest...), followed by the light upscale redo:
renaissance painting of a female alchemist working in her laboratory --aspect 3:4
Much better.
Post-Processing
There's almost always some stray line or dot marring the image. There's nothing wrong with popping it into an editor to spend five minutes cleaning it up manually. You don't need very much skill - most full-featured editors have a "repair" or "magic" brush that you just need to drag over the scraggly bit and it will auto-fill its best guess.
Failure
For every ooh-aah result I post, there were at least a dozen meh-to-awful ones. There's a reason I pay for the unlimited personal plan :)
Gallery
Please enjoy a variety of my results with the exact prompts, some of which may be very unrelated to the actual result. These are all raw unedited outputs.
a Greek goddess of literacy weaving upon a loom --aspect 3:4
portrait of an African sun goddess, rococo painting by Lampi the Elder --aspect 3:4
glass orbs in soft colors, tilt shift photography --aspect 16:9
羊
(this is the Chinese character for sheep)
cute kaomoji papercraft --aspect 16:9
cute kaomoji stickers --aspect 16:9
sky palace video game level --aspect 16:9
a jellyfish made out of galaxies --aspect 3:4
cute illustration of glitch twitter icon
cute illustration of holographic glitch bubbles --no eyes face --aspect 16:4
cute illustration of glitch cupcakes --aspect 16:4
the source code of reality --aspect 3:4
art nouveau illustrated poster of a cathedral in the sunset, pink, purple, orange --aspect 16:9
art nouveau illustrated poster of a beautiful girl with brown skin and blood-red long straight hair, wearing a red and yellow gown with armenian embroidery, in the style of Alphonse Mucha --aspect 3:4
art nouveau cozy cottage in the forest, lit from within, in the style of Ken Sugimori --aspect 16:9
diamonds made of galaxies, raytraced --aspect 16:9
cute user interface for a video game about programming in soft colors --aspect 16:9
underwater video game level --aspect 16:9
portrait of Classical Chinese queen of galaxies --aspect 3:4
