A genre tag, as far as the model is concerned, is a coordinate. It narrows the territory it is willing to wander in. It does not guarantee that the resulting track sounds like the specific record sitting in your head, and treating it as a guarantee is the fastest way to be disappointed by every generation you ask for.
Tags are seeds, not contracts
When you pick "synthwave" from the dropdown, you are telling the model: the territory I want is somewhere that has saw-wave basslines, gated drums, and a shimmer pad somewhere. You are not telling it which subregion of synthwave — Carpenter-tense, Drive- soundtrack-bright, vaporwave-melted — you live in. To get there, the prompt and mood have to do work the genre tag cannot.
Pair every tag with a counter-tag
A surprisingly reliable trick: give the model one genre tag and one mood that does not naturally belong to it. "Synthwave + tender." "Trap + nostalgic." "Folk + paranoid." The model has to negotiate between the two, and the negotiation is where the interesting generations live. Two tags that agree with each other produce the average; two tags in mild conflict produce a take.
When to drop the tag entirely
If the prompt is already specific enough — "a song that sounds like the moment after a fire alarm stops in an empty office building" — the genre tag may actually fight you. The model will try to fit a genre on top of an image that does not need one. In those cases, leave the genre dropdown alone and let the prompt carry the whole weight.