Why doesn't "masterpiece" or "high quality" work?
In our measurements on ACE-Step 1.5, quality words like "masterpiece" and "viral hit" moved the sound about as much as any random edit does, roughly 0.66 to 0.78 on our spectral difference scale, which is not evidence they did anything specific. Confirmed production words moved a measurable property in the direction their name promises: "bright" raised the treble by 11%, "big reverb" widened the stereo image by 68%, "warm analog" cut treble by 22%. Quality words have no direction to check, so the model has nothing concrete to act on.
Which words actually change how polished a track sounds?
Use words tied to a specific, audible effect rather than a vague judgement:
- warm analog - rounds off harsh top end, good for a softer overall feel
- bright - raises treble presence, good for clarity and punch
- big reverb - widens the stereo image, makes a track feel roomier
- delay - also widens the image, adds a sense of depth without changing tone as much as reverb
- dry, close-mic vocal - pulls the sound forward and tight, the opposite of big reverb
Pick one direction (warm or bright, wide or tight) rather than stacking opposites.
How many descriptors should a pop prompt have?
Aim for 4 to 7 descriptors covering drums, bass, lead or vocal, mix and mood, plus a tempo. For pop specifically, most tracks that sound finished sit in the 90 to 130 BPM range, with 100 to 125 BPM most common. A four-chord loop in C major, G major or A minor is a safe harmonic default if you are not specifying chords. Example: "pop, punchy drums, warm bass, bright vocal, wide mix, uplifting mood, 118 BPM" covers all five without overloading the prompt.
Does mastering after generation help more than prompt words?
Yes, for overall polish, mastering (the final pass that balances loudness, EQ and stereo width across a whole track) does more than any single prompt word can. Lyro Music mixes and masters every track it generates, whether it starts from a description, a vocal recording or a beat, so the finishing pass is handled after generation rather than by chasing quality words in the prompt itself.
Common mistakes
- Using "masterpiece" or "high quality" as if they were style tags: replace them with a named trait like "bright" or "warm analog".
- Stacking contradictory polish words such as "warm and bright and dry and reverb-heavy" in one prompt: pick one direction per pair and drop the rest.
- Adding percentages or dB values like "bass +12dB": these read as placebo in our tests, describe the effect in words instead, such as "deep bass".
- Leaving out a tempo when quality is the goal: an unclear BPM often produces a looser, less controlled-sounding mix, so give a number.
- Expecting one word to fix a muddy or thin mix: combine a drum descriptor, a bass descriptor and a mix descriptor together rather than relying on a single adjective.