What a seed fixes
A seed fixes the random initialisation of a generation. Given identical inputs — same source image, same prompt or template, same model version — a fixed seed should reproduce a similar result. It does not make the pipeline deterministic, and services often change models without notice.
What a seed does not fix
It does not fix an unsuitable source photo, a mismatched template, or a motion request the model cannot express. Seeding a bad setup just produces a reproducible bad clip.
Using seeds to isolate variables
- Fix the seed, vary the prompt or template — you are testing the instruction.
- Vary the seed, keep everything else — you are testing run-to-run variance.
- Fix both and change the source photo — you are testing the input.
Model versions move under you
Adult generators update their backends regularly. A seed that worked in March may behave differently in June. Record the model label and date alongside the seed so a future failure is diagnosable rather than mysterious.
The practical protocol
Once you have a result you like, record source, template or prompt, seed, model label and date. That five-field manifest is what turns a lucky generation into a repeatable process.