Midjourney

Midjourney's --seed explained, with examples

By

Yubin

Updated

You'll notice if you repeat the same prompt in Midjourney you'll get fairly random results. If you want more consistency in your generations, you can set a fixed seed value.

Seeds are numbers used to initialize generations. If you don't specify a seed, Midjourney will randomly generate one for you. Every image (and every 4-quadrant image set) has a seed.

Large or small, the actual numbers don't mean anything. Think of them as bookmark for particular stylistic or thematic outcomes; we use numbers because they're convenient and there are a lot of them.

Note: Midjourney does not produce identical results with the same seed, it only produces similar results. If you are looking for a way to generate the exact same image with a given prompt and given seed, you should check out Stable Diffusion instead.

Modifying Images

I'm going to change the seed while keeping images the same.

Modify the settings

Here I use the same prompt and same seed, but different aspect ratios. Check out the similarities across corresponding quadrants:

Landscape Aspect Ratio

light watercolor, girl and robot on beach, white background, few details, dreamy Studio Ghibli --seed 101

Portrait Aspect Ratio

light watercolor, girl and robot on beach, white background, few details, dreamy Studio Ghibli --seed 101 --ar 3:2

Here's another aspect ratio. Again, notice the similarities between corresponding quadrants:

Original

light watercolor, girl and robot on beach, white background, few details, dreamy Studio Ghibli --seed 101

Portrait Aspect Ratio

light watercolor, girl and robot on beach, white background, few details, dreamy Studio Ghibli --seed 101 --ar 2:3

--stylize Command

Here's me using the stylize command. The differences are there, but they're quite small. To take full advantage of stylize, you should change to Midjourney Version 3, you'll get crazier results.

light watercolor, girl and robot on beach, white background, few details, dreamy Studio Ghibli --seed 101
light watercolor, girl and robot on beach, white background, few details, dreamy Studio Ghibli --seed 101 --stylize 1000

Modify the prompt

If you modify the prompt you introduce new variables into the generation. The more drastically you modify the prompt, the more randomness you introduce. Let's see the results of smaller and larger prompt modifications:

Make the Robot Pink

The first time I did this I was surprised by the consistency that locking the seed value gave me.

Check out the second image in both quadrants here:

light watercolor, girl and robot on beach, white background, few details, dreamy Studio Ghibli --seed 101
light watercolor, girl and pink robot on beach, white background, few details, dreamy Studio Ghibli --seed 101

Swap a Subject

I liked how the compositions and colors stayed they same, even with larger subject modifications:

light watercolor, girl and robot on beach, white background, few details, dreamy Studio Ghibli --seed 101
light watercolor, boy and robot on beach, white background, few details, dreamy Studio Ghibli --seed 101
light watercolor, girl and gundam on beach, white background, few details, dreamy Studio Ghibli --seed 101

New Environment

This is definitely a bigger change. The last generation looks very different from the first:

light watercolor, girl and robot on beach, white background, few details, dreamy Studio Ghibli --seed 101
light watercolor, girl and robot on lake shore, white background, few details, dreamy Studio Ghibli --seed 101
light watercolor, girl and robot on lunar surface, white background, few details, dreamy Studio Ghibli --seed 101

Change the Artistic Style

This one is a much more drastic change. Still, you can see which elements are preserved

light watercolor, girl and robot on beach, white background, few details, dreamy Studio Ghibli --seed 101
light watercolor, girl and robot on beach, white background, few details, dreamy by John Singer Sargent --seed 101

Change the Medium

Another interesting one: I quite like the gouache style. You can see it had an effect on the way Midjourney generated the subject matter as well:

light watercolor, girl and robot on beach, white background, few details, dreamy Studio Ghibli --seed 101
gouache, girl and robot on beach, white background, few details, dreamy Studio Ghibli --seed 101

Re-order the prompt

Did you know that prompt order matters? Words towards the front have priority over words in the back. Let's see what happens when we keep the seed the same while re-ordering the prompt:

Using "--sameseed"

I mentioned that one seed corresponds to a 4-quadrant image set.

Will closer seeds produce closer images?

For example: will images created with --seed 10 and --seed 11 look more similar to each other than an image created with --seed 1000000?

Nope! A seed is just a way to get the same starting data for image generation, but two different seeds will have totally different starting data regardless of how close they are to each other.

Finding the seed of an existing image

Let's say you didn't specify the seed. How do you find it after the image has already been generated? React to any image you've generated with the envelope emoji: ✉️, and the Midjourney bot will send you the seed. More details here.

Unfortunately, you cannot find out the seed of other peoples' images. Your best bet is to ask them for it nicely.

That's the article! Also check out:

FREE RESOURCE
The Big Book of Prompts: Midjourney V5 Lost Worlds

Endless hours and thousands of prompts of experimentation later... I’m happy to release the Midjourney V5 book:

  • 94-pages, 200+ prompts, and time saving techniques
  • Focus on fantasy and worldbuilding
  • Create high quality assets to use in games, concept art, and hobbies

That's the article! Also check out:

FREE RESOURCE
SDXL Prompt Magic

A handbook that helps you improve your SDXL results, fast. Includes:

  • Easy step-by-step instructions
  • My favorite SDXL ComfyUI workflow
  • Recommendations for SDXL models, LoRAs & upscalers
  • Realistic and stylized/anime prompt examples

2 thoughts on “Midjourney's --seed explained, with examples”

Leave a Comment