seed
A seed is a starting number that determines what 'randomness' the generation begins with, most often for creating an image; the same seed with the same model, prompt, and settings typically yields the same result.
In plain terms
When you ask an AI image generator for the same thing twice, you get two different images. The reason is that the machine works from a random starting point each time. The seed is the serial number of this starting point: a simple integer, for example 123456.
It is like shuffling a deck of cards. If someone specifies exactly which shuffle was used, the game can be replayed card by card. The seed is the shuffle number: if you remember it, you get the same 'deal' back.
Why it helps: if you like an image, fixing the seed lets you keep its foundation and refine only the description. This way you see what changed with the single word you edited, instead of getting a completely new image with every attempt.
A seed is not an image description nor a style setting. By itself, it says nothing about what will appear in the image; it only determines the outcome together with the prompt, the model, and other settings.
How it works
- Most programs do not use true randomness: a formula generates a sequence that appears random, and this sequence always starts from the seed.
- The most widely used so-called diffusion image generators today create a starting 'noise image' from this number sequence, which looks like a snowy TV screen.
- The model cleans this noise step by step based on your prompt until the image emerges.
- The same seed produces the same noise, so with identical prompts, models, and settings, the final result will be the same, with only minor variations on different machines.
- If you do not provide a seed, the program draws one itself, and many tools write this into the data of the finished image so you can retrieve it later.
An everyday example
You generate an image of a red cat on a windowsill, and the generator nailed the scene, except the background is too dark. You copy the image's seed, re-enter it, and add "sunny morning" to the prompt: you will likely get a similarly structured, brighter image, although the cat's details may also change somewhat.
Who is affected
A common misconception
Limits and risks
- A seed cannot be transferred from one service to another: the same number means something different elsewhere.
- After a model update, the old seed may not produce the same image, so save important images, not just the number.
- Not all tools allow specifying or showing the seed; chat-based image generators often hide it.
- For text models, the seed at most increases reproducibility, but providers generally do not guarantee identical responses letter by letter.
Similar, but not the same
- prompt – The prompt tells you what you want to see, while the seed determines which of the many possible implementations you receive.
- temperature – Temperature controls how much randomness enters the output, while the seed determines exactly which randomness occurs.
- watermark – A watermark is a hidden mark embedded in the final content to verify its origin. The seed is the starting setting for generation: it cannot be determined from the image itself, but can be read from data saved alongside the file if the tool wrote it there.
Related concepts
generative AI, pseudo-random number generator, reproducibility
fixed seed, random seed
prompt, diffusion model, temperature, image generator, sampling
A brief history
The seed is an old concept in computer random number generation: as early as the late 1940s, John von Neumann proposed a method that generates a seemingly random sequence of numbers from a starting number. The general public became familiar with it through open image generators released in 2022, mainly Stable Diffusion, where the seed became a user-adjustable field.
Where things stand
By the end of 2026, a seed is a standard setting in image, video, and music generation tools, available almost everywhere on advanced interfaces and developer access points. Simple chat interfaces, however, often hide it, and providers rarely guarantee full reproducibility.
In the news
- DIMSZ, the Hungarian data-driven marketing association, continues with a new boardNov 7, 2026 · Marketing
- Survey: four in five industrial players face skills shortages; two-thirds of AI users also use personal devicesNov 4, 2026 · Ipar
- Budapest to become a starting point for the autonomous factory eraNov 4, 2026 · Ipar
- OpenAI strikes deals with Magnite and PubMatic to fill unused ChatGPT ad spaceNov 2, 2026 · Marketing
- Clark raises $11M seed round for autonomous AI auditorsNov 2, 2026 · Money
- Robotics and industrial automation startups raise major funding: Bleu, Foundry, NeuronNov 2, 2026 · Ipar
- Hungarian startup Kodesage closes $6.6 million seed roundNov 1, 2026 · Code
- Chinese robots and open AI models could reshape Hungarian jobs tooOct 30, 2026 · Ipar
Also known as
seed · random seed · random seed · mag · intrinsic value · initial value · seed value · seed number
Sources
This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.