In short
- A seed sets the random noise an image generator starts from. The same prompt, seed and settings can reproduce a picture, and a different seed gives a different one.
- Reproducing an image takes more than the seed: the same model version, sampler, step count, guidance and resolution too.
- Even then, different hardware and software versions can produce slightly different pixels.
- A style lock pins a seed together with a frozen style embedding, so a series of prompts keeps one consistent look.
- Logging the settings next to every image makes results reproducible, reviewable and auditable.
What a seed does
Many image generators, including diffusion models, start from random noise and refine it step by step into a picture guided by your prompt. The seed is the number that decides which noise you start with. Same seed, same starting noise.
Keep the seed and change the prompt slightly, and you often get a similar composition with different details. Keep the prompt and change the seed, and you get a different image of the same idea. That makes the seed a useful control when you want to find out what a change in wording actually did.
What has to match to get the same image
- The model: the same weights at the same version. Record a checksum of the files.
- The prompt: the exact text, including any negative prompt and other conditioning.
- The seed.
- The sampler and the number of steps.
- Guidance strength, the setting that controls how closely the image follows the prompt.
- Resolution and aspect ratio.
- Any style lock or other conditioning.
- The software: library versions and numeric precision settings.
Why results can still drift
Setting a seed doesn't guarantee identical output everywhere. PyTorch's documentation says results are not guaranteed to be reproducible across releases, individual commits or platforms, and that results can differ between CPU and GPU even with identical seeds. Some GPU operations are nondeterministic unless you ask for deterministic implementations, which are often slower.
In practice, to reproduce an image exactly, use the same model files, the same software versions and the same kind of hardware, and switch on your framework's deterministic settings where they exist. To reproduce an image's look, the seed and settings are usually enough.
Style locks: keeping a series consistent
A seed alone doesn't hold a style. Change the prompt and the look can change with it. A style lock, as we use the term for Synas Image 001, pins a seed together with a frozen style embedding, so a series of prompts renders in one consistent look. Same lock, same style, every time.
That is useful for a campaign, a storyboard or a dataset, anywhere images have to belong together.
What to log with every image
Save the settings alongside the file. The field names differ from tool to tool, but a record like this covers the essentials:
model, weights checksum, prompt, negative prompt, seed, sampler, steps, guidance, width and height, style lock, software versions, date.
With that record, a reviewer can regenerate what they are approving, and you can tell later exactly how an image was made.
Why it matters
- Fair comparisons. Change one thing at a time and keep the seed fixed, and you can see what that change did.
- Regression tests. Re-run a fixed set of prompts and seeds after a model update to see what changed.
- Review. Anyone who approves an image can regenerate it.
- Audit trails. A log shows how an image was made, which complements the signed provenance others can verify.
Making, labeling and checking
A generation log tells you how you made an image. Content credentials let others verify claims about it. A detector, such as Synas Detect 001, can check files whose history you don't know. Together they cover making, labeling and checking images.
Sources
- PyTorch documentation: Reproducibility. Results aren't guaranteed identical across releases, platforms, or between CPU and GPU