Reproducibility

What reproduces is the process and the decisions, not the waveform.

Related

DiffusionSample Rate

The problem

The same prompt and the same seed should give the same result, except in AI music they often do not. Sampling is stochastic, models update, server-side preprocessing is invisible, and changing one reference file changes everything. Chasing a bit-identical waveform is usually wasted time, but a result that cannot be reproduced at all cannot be revised or defended.

Why it fails

Three reasons. The generation process samples randomly, and a seed fixes only part of it. Models and inference services are upgraded, so the same name now hides different weights. Inputs include more than the prompt, namely reference audio, duration and hidden settings, and changing any one forks the result. What can be reproduced is the process and the decisions, not the waveform. Separating the two is the whole method.

Recipe and testing

A useful recipe lists tool and version, model, prompt verbatim, negative prompt, seed, duration, tempo, key, hash and licence of the reference audio, input MIDI or audio, number of candidates, selection criteria, post chain, and a line of known unknowns. Test once per project, run the same recipe three times and compare duration, section boundaries, spectral difference and a blind listen, then write down the tolerance, for example same structure with timbral drift. Prefer tools that expose seeds for client work; for archives, freeze the output rather than trusting the service. If three runs differ wildly, drop that tool from the candidate list for client work.

What to listen for

In the twenty-six minute Toccata, how small deviations accumulate into structural differences over a long file. In Bach's Toccata in D major, free passages whose decisions cannot be recovered from parameters. In the 138 BPM funk loop, the part that is determined, since pattern and tempo can be repeated but timbre may not. In Scriabin's Second Sonata, rubato that no recipe can exhaust.

Sources

• 生成式模型的随机性与版本漂移，依机器学习公开资料核对，未抄原文
• 复现测试的容差与记录方式，依工程实践通识归纳，未抄文本

Used for fact-checking only; all prose is original.