prompting.

paste it into the part’s description box in the plug-in. the starter chips there are built the same way.

stems takes the parts you already have, summed, and writes one part to fit. the text it reads looks exactly like this:

plus the lyrics, on their own. you control the caption, the lyrics and the context audio. everything else is fixed — the plug-in never sends a tempo or a key, because this model never saw one; it locks both from the audio itself.

lower-case. two genres. one or two character words. the instrument word. an optional placement. a full stop. anything else is off-distribution — the model never saw prose like “a tight drum kit locked to the groove”. never leave it empty; the plug-in fills an empty box with a real caption for you.

genres — two, comma-separated. the strongest pairs, in the model’s order:

emo rap, cloud rap indie rock, indie pop indie pop, bedroom pop trap, hip hop midwest emo, emo indie rock, alternative rock post-hardcore, screamo emo, indie rock hyperpop, electropop indie rock, bedroom pop hyperpop, digicore indie folk, singer-songwriter emo, pop punk hip hop, trap indie rock, grunge indie pop, funk indie pop, dream pop shoegaze, dream pop

the corpus is emo / indie / bedroom-pop / rap heavy. “jazz, bossa nova” or “techno, house” are weak or absent — expect a drift toward the nearest genre it knows.

placement (optional): loud and forward sitting low in the mix dropping out in places with long gaps between phrases entering and dropping out

good:

• indie rock, grunge. clean, active bass.
• emo rap, cloud rap. bright, sparse drums, sitting low in the mix.
• indie pop, bedroom pop. smooth, plucky lead melody, dropping out in places.
• midwest emo, emo. warm, slow and sustained vocals, loud and forward.
• post-hardcore, screamo. harsh, dynamic vocals.

bad:

• “a lead vocal locked to the melody and phrasing” — prose. not a sentence it knows.
• “balanced vocals.” — no genre. almost nothing in training looks like this.
• “Vocals, female, pop, 120bpm” — wrong grammar, and tempo doesn’t belong in words.
• “synth lead melody” — there is no “synth” in its vocabulary. melody was learned as guitar-or-unspecified; steer timbre with distorted / clean / smooth.

• one lyric line per text line. no [verse] / [chorus] tags, no timestamps, no blank-line sections — training lyrics were untagged lists. tags aren’t fatal, just noise.
• size to the window. about 6–12 lines for 48 s. a 30 s take wants ~5–8; two minutes can carry 20+. too few → long wordless gaps or repeats; far too many → it rushes and drops the tail.
• repetition is natural. a chorus line 2–3× looks like real lyrics.
• english. that’s what it learned. other languages get sung with an accent at best.
• lyrics are required for vocals — an empty lyric is how the model says “no vocals here”, so there is no wordless-vocal mode.
• keep the vocal out of the context. it was trained on mix-minus-target; a vocal already in there makes it sing around it strangely.

• drums + bass for a melody. drums + bass + guitar for vocals. a full instrumental works. one lone stem works too, just weaker.
• 10–120 seconds. 48 s is the sweet spot (the training window). the generated part matches the context length exactly — the plug-in keeps it to whole bars.
• clean stems beat a muddy mix — it heard separated stems in training.
• tempo and key are read from the audio. there is nothing to type.

• steps: 64 for a take you’ll keep; 24–32 to audition. quality is flat above ~48.
• seed: same seed ≈ same take (the gpu adds a little nondeterminism). seed variance is large — new take in the plug-in rolls it; refine keeps it and changes the words.
• want a different feel? swap character words (sparse ↔ busy, clean ↔ distorted) or placement before touching genre. genre pulls hardest and can change the whole part.

• drums — punchy, tight for modern / rap; washy, dark for emo / indie. sparse genuinely thins the pattern; busy adds fills.
• bass — active / plucky for movement; sustained / deep for pad-style roots; gritty adds drive.
• melody — the instruction says guitar, and it leans guitar-ish whatever you write. distorted / clean / smooth steer timbre; busy / sustained steer density.
• vocals — rapid and rhythmic pushes toward rap; slow and sustained toward sung; harsh toward screamed; warm, even is the safest sung default. with long gaps between phrases leaves room for the instrumental.