Seeds get described as though they store an image — find the magic number, get the picture back. They do not. A seed sets where the process starts, and the image that emerges depends just as much on everything else you configured. Understanding which is which turns the seed from a lottery ticket into the most useful experimental control you have.
What it actually is
Image generation begins with random noise, which the model progressively denoises into a picture. The seed initialises the random number generator that produces that noise.
Same seed, same starting noise. Combined with identical settings, that gives an identical result — reproducibility, not storage.
Where the seed enters
The seed touches only the first step. Everything after is deterministic given that noise plus your settings — which is exactly why reproduction works, and why it breaks the moment any setting changes.
💡 The seed is a coordinate, not a container
Think of your settings as defining a space of possible images, and the seed as picking a point in it. The number does not describe the picture; the space does.
Change the prompt and you have defined a different space, so the same coordinate lands somewhere unrelated. This is the whole reason seeds are not portable between setups — a coordinate means nothing without the space it indexes.
What must match to reproduce an image
All of it. Any single difference produces a different image, sometimes subtly and sometimes completely.
| Must match | Effect if it differs |
|---|---|
| Model and version | Completely different image |
| Prompt, exactly | Different image — punctuation counts |
| Negative prompt | Different image |
| Sampler | Different image |
| Step count | Different image |
| Guidance (CFG) scale | Different image |
| Resolution | Completely different — noise shape changes |
| VAE | Same composition, different colour and detail |
| LoRAs and weights | Different image |
Resolution is the one that surprises people. The noise tensor's dimensions come from the target resolution, so generating 512×768 instead of 512×512 does not crop or extend the image — it produces an entirely different starting noise and therefore an unrelated picture.
⚠️ Identical settings can still differ
Even with everything matched, reproduction is not guaranteed bit-for-bit across environments. GPU and CPU random number generators differ, floating-point operations are not always associative in parallel execution, and library versions change kernel implementations.
In practice the same seed on different hardware usually gives a very similar image rather than an identical one — same composition, small differences in detail. If you need exact reproduction, pin the hardware and the library versions, not just the parameters.
Using seeds as a control variable
This is where the seed earns its keep, and it is the standard method of any experiment: hold everything constant except the one thing you are testing.
Both halves matter. Judging a prompt change on a single random seed tells you almost nothing — you cannot separate the effect of your edit from the effect of a different starting point. Equally, judging a prompt from one seed tells you nothing about whether the prompt is reliable or you got lucky.
The practical workflow most experienced users converge on:
- Explore with random seeds until a composition appears that is roughly right.
- Lock that seed and refine the prompt against it, one change at a time.
- Re-test the finished prompt across several fresh seeds to confirm it works generally rather than on that one.
- Record the seed with the final output, so the result can be regenerated or extended later.
Nearby seeds and variation strength
A common misconception worth clearing up: seed 1234 and seed 1235 are not similar. A pseudo-random generator is designed to produce uncorrelated output from adjacent inputs, so consecutive seeds give completely unrelated noise. Stepping through seeds is a way to sample the space broadly, not to make small adjustments.
For genuinely small variations, tools offer a variation or subseed strength, which blends the original noise with a second noise tensor:
That is the correct tool for "like this, but slightly different". Incrementing the seed is not.
Where seeds live
Most tools write the seed into the PNG's text chunks alongside the prompt, sampler and model hash — which means an unmodified generated file typically contains everything required to reproduce it.
Two consequences follow. Sharing a generated PNG usually shares your full prompt and settings, whether or not you intended to — covered in AI image generators hide your prompt in the file. And converting to WebP or JPEG, or uploading to most platforms, destroys this — so your own reproducibility depends on keeping the original file.
🚨 Keep the PNG or lose the seed
If your archive is the published WebP, the generation parameters are gone. You cannot regenerate the image, produce a variation, or make a higher-resolution version of the same composition.
Keep the original PNG as a master, and publish converted derivatives — the same master-and-derivative workflow that saves you bandwidth also preserves your ability to go back.
Checking what is embedded in your generated images?
Inspect and strip image metadata entirely in your browser — nothing is uploaded to a server.
Open EXIF Remover →Summary
- A seed initialises the starting noise, nothing more.
- It is a coordinate, not a container. The settings define the space it indexes.
- Everything must match to reproduce — model, prompt, sampler, steps, CFG, resolution, VAE.
- Resolution changes the noise shape, so it gives an unrelated image, not a crop.
- Hardware and library differences can prevent bit-exact reproduction.
- Fix the seed to test prompts; vary it to test prompts' reliability.
- Adjacent seeds are unrelated. Use variation strength for small changes.
- Seeds live in PNG metadata and die on conversion or upload.
Frequently Asked Questions
What does a seed do in AI image generation?
It initialises the random number generator that produces the starting noise the model denoises into an image. Because that noise is the starting point of the whole process, the same seed with identical settings reproduces the same result, and a different seed produces a different image from the same prompt.
Does the same seed always produce the same image?
Only if everything else matches exactly — the same model, prompt, negative prompt, sampler, step count, guidance scale and resolution. Change any one and the result differs. Even with all settings identical, different hardware or library versions can produce small differences because floating-point operations are not always bit-identical.
Can I use a seed from someone else's image?
Only with their exact setup. A seed is meaningless without the same model, prompt and settings, so the same number on a different model produces a completely unrelated image. Seeds are only portable alongside the full generation parameters.
How do I get variations of an image I like?
Keep the seed fixed and change one thing at a time to see the effect of a prompt or setting. To get similar but different compositions, use nearby seeds or a variation strength setting if your tool offers one, which blends the original noise with a small amount of new noise.
Where is the seed stored?
Most tools write it into the PNG metadata alongside the prompt and other settings, so an unmodified generated file usually contains everything needed to reproduce it. Converting the image to another format or uploading it to a platform typically strips that metadata.