Cross-platformParameters & settingsPrompting basics

Same Prompt, Different Result: Seeds and Reproducibility Explained

By Marco Cavazzana · Updated · 5 min read
ShareEmail
Short answer

The same prompt gives a different result because each generation starts from random noise, chosen by a number called the seed. A new seed means a new starting point and a new image or clip. Some tools let you fix the seed: Midjourney with --seed, and Runway Gen-4 with a Fixed seed setting. That gives similar results with the same prompt and settings on the same model. Even then, results aren't guaranteed to match across sessions or model updates. For consistency, use references and keep a full record of each generation.

You found a great result, typed the same prompt again, and got something completely different. That isn't a bug. It's how diffusion-based image and video models work. This guide explains what's going on, which tools let you control it, and how to get as close as possible to repeatable results.

Why the same prompt gives a different result

Image and video models don't look up a picture that matches your words. They start from a field of random noise and refine it, step by step, towards something that fits the prompt. Different noise leads to a different composition, pose, lighting and detail, even with identical words. Midjourney's documentation compares the seed to "the random noise you see on a TV screen": it is the first step before the image takes shape.

The seed is simply the number that decides that starting noise. Leave it unset and the tool picks a new random one every time, which is why you get variety. Set it, and you lock the starting point.

What else changes the result

The seed is only one input. The result also depends on:

  • The model and its version. A new version reads the same prompt and seed differently. Midjourney's docs say seeds have "the least impact on the final image" next to the prompt, model version and settings.
  • Settings and parameters: aspect ratio, stylise, motion, quality, resolution, duration. Change one and the result changes.
  • References: any image, style or character reference attached to the prompt.
  • Speed modes. Midjourney says not to use Turbo mode with seeds, because seed locking isn't reliable there.
  • Silent updates. Platforms update models, and sometimes switch which model a default setting uses. Midjourney warns that if you "take a break and come back later, your seed may no longer work as expected".

Which tools let you set a seed

ToolSeed controlWhat the vendor says
MidjourneyAdd --seed and a whole number from 0 to 4294967295. Copy any image's seed from its options menuSeeds are "99% identical in V8.X", but meant for testing, not for saving styles or long-term reproduction
Runway Gen-4 and Gen-4 TurboFixed seed slider in the settings; paste a previous output's seedGives "similar style and movement", not an identical clip. Off by default
Google Flow, ChatGPT, Gemini, KlingWe couldn't find a seed control in their consumer help pagesCheck the settings panel in your version. Use references for consistency instead

If you build through an API, check that model's documentation for a seed parameter, since API options can differ from the app.

What a seed is good for

The best use of a fixed seed is as a controlled experiment. Keep the seed and settings the same, change one word in the prompt, and you can see what that word does without the noise of a fresh random start. Midjourney recommends exactly this, calling a locked seed "like having a control in an experiment". It's also useful for making small variations of a result you like.

What a seed is not good for: keeping a character, product or style consistent across different prompts. Midjourney says seeds "can't capture or bookmark a specific style, character, or appearance", and recommends style references, personalisation and its reference tools. On video platforms the same applies: reference features such as Kling Omni Reference, Flow Ingredients or Runway References do this job. Our consistent characters guide compares them.

Seeds in video

Video adds a second source of variation: motion over time. A fixed seed fixes the starting noise, but every frame after that is still decided by the model, so small differences can grow over a clip. Runway describes its Fixed seed as giving "similar motion" and "similar style and movement", which is the honest promise. For video, the strongest lever for repeatable results is a fixed first frame (and, where available, a fixed end frame), not the seed.

How to get repeatable results

  1. Record everything: the exact prompt, model and version, seed (if the tool shows it), every setting, and the reference images.
  2. Pin the model version for the whole project instead of following the default, where the tool allows it.
  3. Start from the output, not the prompt. If you need "the same again, slightly different", use the platform's variation, remix or edit tools on the result you liked. That is more reliable than re-rolling the prompt.
  4. For video, start from an approved still. Image-to-video with the same first frame gets you much closer between runs than text-to-video.
  5. Use references for identity and style, and seeds only for testing.
  6. Accept that some drift is normal. Treat a match as "close enough for the client", not pixel-perfect, unless you still have the exact model, seed and settings.

For Midjourney specifically, see our Midjourney seeds guide.

Try it

In Midjourney, run a simple prompt with --seed 1234. Run it again with the same seed and settings: the results should be nearly identical. Now change one word, say "morning light" to "evening light", and keep the seed. The difference you see is that word's effect alone. Then remove the seed and run the original again to see how much a fresh random start changes things.

Keeping the recipe

Step 1 is the hard part, because few tools keep all of it visible in one place, and it's gone the moment you lose track of a session. Dreamdrive, a free Chrome extension, saves each image and video automatically as you make it, with its prompt, model and settings, so the recipe behind a result you liked is still there when you need to get close to it again.

Related: what happens when a model is retired.

Frequently asked questions

Why does the same AI prompt give different results?
Each generation starts from random noise chosen by a seed number. A new seed gives a new starting point, so the image or clip changes even with identical words. The model version, settings and references also affect the result.
What is a seed in AI image generation?
A seed is the number that sets the starting noise a model refines into an image or video. Fixing it, for example with Midjourney's --seed parameter, lets you repeat or compare results with the same prompt, model and settings.
Does Runway have a seed setting?
Yes. Runway's Gen-4 and Gen-4 Turbo have a Fixed seed setting. Paste the seed of a previous output to get generations with similar style and movement. It is off by default.
Can I recreate an AI image exactly?
Only if you have the same model version, prompt, seed and settings, and even then Midjourney warns seeds may not behave the same across sessions. For a close match, start from the original output with variation or edit tools, or use it as a reference.

Nothing you makeis lost.

Free · No card · Read-only · 2-minute setup

DreamdriveFree plan · Chrome
Add to Chrome — it’s free