Seeds and Reproducibility

Log the seed and every parameter, but do not expect two runs to match exactly.

Related

DAWMIDI ProtocolStems

The problem

You tune a result you like, ask for another version the next day with the same settings copied across, and get something else. The seed is only one variable among many, and treating it as the single switch makes any later review collapse. Worse, some services offer no seed at all, only a regenerate button.

How it works

A result is decided by model version, weights, sampling settings, prompt text, reference audio and processing order together. Once a service updates in the background, an old seed points into a new distribution. Even with all of that frozen, floating-point arithmetic on different cards can yield small differences that repeat after repeat turns audible. Only symbolic output stays stable, because a written note stays written. The same prompt placed in batches of different length is also cut at different points and returns different results.

Practice

Keep a log for every run, recording the date, tool and version, model name, seed, steps, guidance strength, duration, the prompt as written, any reference file and the output filename. Store the log beside the take you accepted so the pair still matches months later, and keep it in plain text beside the project file so a collaborator can read it. For long-term reproduction, export the accepted result as stems or MIDI and turn the generation into an editable asset, rather than staking reproducibility on the service. When regenerating, move one variable at a time and keep a control version, so that even without an exact repeat you can say which step changed the outcome.

What to listen for

In the aria's return, the difference when the same material comes back; in the Passacaglia, the identity kept through a repeating bass; on the hip-hop kit, the small wobble between one pass of the loop and the next; in the folk tune's repeats, how the second pass differs. After hearing these, an inexact match stops looking like failure, since an exact match is the exception and a documented approach is the rule.

Sources

• 生成参数与随机种子的作用，依公开模型文档归纳，未抄原文
• 版本更新导致结果漂移的现象，依制作实践核对，未抄文本

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