Midjourney Seeds: How to Get Repeatable Results

What a seed actually does

Midjourney builds every image from a field of random visual noise, then gradually shapes that noise into something that matches your prompt. The seed is the number used to generate that starting noise.

When you don’t set one, Midjourney picks a random seed for each job. That’s why the same prompt gives a different grid every time. When you set the seed yourself, the starting noise is the same, so the result is far more predictable.

The syntax is a parameter at the end of your prompt:

a glass teapot on a wooden table, morning sunlight, steam rising --ar 4:5 --seed 1234

The seed must be a whole number. You can choose any number in Midjourney’s accepted range; there’s nothing special about one number over another.

What stays the same and what doesn’t

A seed only gives repeatable results when everything else is also the same. This table shows what to expect.

The key takeaway: a seed fixes the starting point, not the content. It isn’t a way to lock a character, face or style.

Using seeds for controlled testing

Seeds are most valuable when you want to understand what each part of your prompt does. With a fixed seed, any difference you see comes from your change, not from randomness.

1. Write the prompt and add a seed, e.g. --seed 777.
2. Run it once as your baseline.
3. Change a single thing: one word, one lighting term or one parameter value.
4. Run again with the same seed and compare side by side.
5. Keep the change if it helped, then test the next one.

For example, to see what --raw really does to your photography prompt:

street portrait of an elderly fisherman, overcast light, 85mm --seed 777

street portrait of an elderly fisherman, overcast light, 85mm --raw --seed 777

This method works well for comparing stylize, chaos and weird values, and for checking whether a negative prompt removes what you intended without damaging the rest.

Finding the seed of an existing image

If you didn’t set a seed, you may still be able to find the one Midjourney used:

• On the website: open the image from your Create or Organize page, use the options button next to the prompt, choose Copy, then Seed.
• On the Midjourney website: the interface changes over time, so check the job’s details and options. If you can’t find it, the simplest habit is to set your own seed whenever you think you may want to repeat something.

Once you have the seed, add it to the original prompt with the original parameters and model to get back to a similar starting point.

Seed workflows for real projects

Beyond one-off tests, seeds are handy in a few everyday situations:

• Client or team reviews. If someone asks "can we see that one again with warmer light?", having the seed means you can change one word and return a closely related image rather than starting from scratch.
• Building a prompt library. Saving a prompt together with its seed and parameters lets you reproduce a reference image later, which is useful when documenting a house style.
• Format versions. When you need the same idea at several sizes, keeping the seed fixed while changing the aspect ratio won’t give identical compositions, but it often keeps the colour and mood closer than a random seed would.
• Pairing with chaos. A fixed seed with --c lets you compare how much variety each chaos value adds, without random noise muddying the comparison.

Keep a short note of seeds that produced good starting points for a project. A seed alone means little, but a seed plus its exact prompt is a reliable record.

The limits of seeds

Seeds are often oversold. Be realistic about what they can do:

• Not a character lock. The same seed with a new scene won’t keep a person’s face or outfit consistent. Use reference images, consistent description and the approaches in our consistent characters guide instead.
• Not a style lock. For a consistent look across many different subjects, a style reference (--sref) is the right tool.
• Not version-proof. A seed from one model version won’t recreate the same image in another.
• Not always pixel-perfect. Identical inputs usually give near-identical results, but small differences can appear. Treat it as "very similar", not guaranteed.
• Not a quality setting. No seed number is better than another. If a seed gave you a lucky image, it was the combination of seed, prompt and settings.

If you already have an image you like and want small changes, Vary (Subtle) on that image is usually more effective than rerunning a seed with an edited prompt.

Doing this faster with Dreamdrive

Dreamdrive keeps the full prompt with every capture, --sref code included, so a look you liked is a search away in your Library.