TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 17—19 · FACTUAL · DURABLE
Seed, temperature, and re-runs
A fixed seed makes a run repeatable, so one changed word shows one change.
The idea
Two settings explain most re-runs. The seed is the number that fixes the random starting point — the exact pattern of noise, or the exact random draw in a text model. A fixed seed makes a run repeatable on the same setup: same model, sampler, software and hardware. Change any of those and the same seed may give a different result. Temperature reshapes the spread the sampler draws from. Low, and the likeliest option wins almost every time. High, and the draw wanders further. Re-rolling without the seed fixed gives a different result you did not steer.
Why it matters
"Same image but change the roof" is impossible without the seed. With it, you change one thing and see one thing change.
See it in the studio
You finally get a massing you like, change "brick" to "stone", and the whole building is different. You re-rolled. Fix the seed, then change the word.
Watch for this
Tools that hide the seed. Many consumer apps do. If you cannot fix it, you cannot run a controlled comparison. Note that before you draw conclusions from two runs.
Try it
Fix a seed. Run a prompt. Change one word and run again. Then unfix the seed and change the same word. Lay the four results in a square and describe what each pair shows.
Prove it
Explain why fixing the seed lets you repeat a result on the same setup while re-rolling only gives a different one, say what else has to stay fixed for the repeat to hold, and describe what turning temperature up does to how spread-out the results get.
How it works
Generation needs random numbers. The seed is the starting value of the generator that supplies them, so the same seed gives the same sequence again. Temperature reshapes the probability spread before the draw — SAMPLING on this strand, THE DRAW on Foundations, seen from the sampler's side. Different tools use different samplers and step counts, so the same seed rarely matches across tools, and PyTorch's own documentation says results are not guaranteed to match across releases, platforms, or CPU and GPU, even with identical seeds. Record the model version and settings with the seed, or the seed alone records nothing.
What this idea builds on
What this idea opens up
Sources
- diffusion-how-ai-paints-from-noise
- ch01
- essential-ai-concepts
- PyTorch, Reproducibility
- Hugging Face Diffusers documentation
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.