Simple is best for fast starts: you describe the idea and generate. Custom is better when structure, your own lyrics, style control, and more predictable output matter.
Simple Mode
Best for sketches, testing ideas, finding a vibe, and fast generations from one short description.
Fast idea testing
Single description prompt
Best for sketches & vibes
Custom Mode
Use it when you want to insert your own lyrics, set style more precisely, and shape the track more intentionally around genre, mood, and vocal delivery.
Insert your own lyrics
Precise style & genre control
Structured sections support
2
How to write a strong prompt
A strong prompt usually combines genre, mood, energy, tempo, instruments, and vocal type. Avoid being too abstract — a few concrete anchors work better than one vague phrase.
Each element of the formula adds precision and reduces randomness in the output
The style field is not just a tag list. Mix genre, sonic texture, delivery, and mood: for example dreamy synth-pop, warm analog pads, punchy drums, intimate female vocal. Do not overload the field: with ten or more descriptors they start to compete.
Mix genre, texture and delivery: e.g. "dreamy synth-pop, warm analog pads, punchy drums"
Add sonic descriptors: "lo-fi, bedroom pop, airy reverb, intimate vocal"
The field takes up to 1000 characters, but there is no need to fill it: 4–7 descriptors are enough, and it is best to start with the main genre
4
Which language to write in
The model sings in many languages, but its musical vocabulary is English: it recognises genres, instruments, and delivery far more reliably from English words. Hence a simple split: the service fields in English, the sung lines in your own language.
Keep the Music Style field in English
«energetic pop-rock, driving guitars, male vocal» is the usual way to describe a style: the model knows genres, instruments and delivery by their English names. The same goes for the Exclude Styles field and for section tags: [Verse], [Chorus] stay in English.
Write the lyrics in your own language
Only the lines between the tags are sung, and the rule does not apply to them. The language of the style field has no effect on them.
Name the vocal language in the style
Add “Russian female vocal, clear Russian diction” to the style. Without it the model sometimes bends foreign words towards English pronunciation.
In Simple mode write in the song language
There the description shapes both the music and the lyrics: an English description gives an English song. Genre words inside the phrase — “indie pop, 92 BPM” — can stay as they are.
5
Good and bad examples
Both examples are about the Music Style field in Custom mode. The “Try it” button drops the prompt straight into the studio.
Good prompt
Melancholic synthwave pop, 95 BPM, deep female vocal, thick analog bass, rainy neon atmosphere, cinematic production
Make a beautiful song. No genre, no mood, no vocal — the model decides everything, differently every time.
6
Lyrics recommendations
If you write your own lyrics, keep a clear structure such as [Verse], [Chorus], [Bridge]. Four to eight lines per verse, two to four per chorus: short rhythmic phrases sit on the music better than long ones. Repeat an identical chorus word for word so the melody sticks; keep the verses different. Put each tag on its own line before what it describes. Instructions about voices and instruments written into the lyrics may be sung aloud: keep the overall sound description in the Music Style field and only short tags in the lyrics.
Use section tags like [Verse], [Chorus], [Bridge] in Custom Mode to control structure
7
Advanced settings
Custom mode hides six settings behind the Advanced Settings button. Their defaults are sensible and you do not need to touch them every time — but each one solves its own problem.
Exclude Styles
This is where you put what must not appear: «trap hi-hats, autotune, country twang». Negations inside the style field itself («no trumpet») are unreliable, which is what this separate field is for. Keep this field in English too.
Vocal gender
A direct choice of male or female voice.
Style weight
How strictly to follow the description. 50% is balanced. Higher — the model follows the tags literally; lower — it takes liberties. Raise it if the style seems ignored.
Creativity control
How unusual the result will be. 50% is normal. Higher — unexpected arrangements; lower — predictable genre classics. For a gift song, keep it below the middle.
Audio weight
Only applies when you have uploaded your own audio — cover, remix, continuation. Higher stays closer to the original, lower is freer. On a plain generation it does nothing.
Duration
A slider from 10 seconds to 6 minutes. If you leave it alone, the network picks the length itself: usually 2–4 minutes, 8 at most.
8
How to choose a model
V6 is a reliable, precise model and the default: it consistently delivers polished genres and styles. V6 Wild is less predictable and more varied, with unexpected, textured and ambitious results. V6 Mini is the fastest.
V6 is the recommended default — start here unless you have a reason to switch
Up to 8 minutes in one generation, and you can set the length manually
Your own voice (Voices) works on every model in the list
9
Common mistakes
A prompt that is too generic and lacks genre or mood
Too many conflicting tags in the style field
Lyrics that are too long and unstructured
Expecting a perfect result from the first generation instead of iterating 2–3 times
An upbeat style over sad lyrics: the mood of the style field and of the lyrics must match
Long instructions in the lyrics: voices and instruments described in full sentences. The model may sing them aloud or perform the text worse
10
What the model does not understand
Some hints that feel obvious are read differently by the model. Here are the main mismatches.
Square brackets define the structure
[Verse], [Chorus], [Bridge], [Intro], [Outro] — only these tell the model which part of the song is which. The words «Verse» or «Chorus» written as plain text are not structure, they are simply sung.
Round brackets: backing vocals and short cues
Round brackets hold backing vocals and echoes: «I will return (return)». Short delivery cues such as (softly) or (whispered), one to three words, are also used, but long phrases like «(guitar plays and rain softly falls)» are best avoided: they may be sung as words. Set the overall sound and instruments in the Music Style field.
Instruments are set in two places
Overall sound goes into the Music Style field. A separate instrumental break goes into the lyrics in square brackets: [Guitar solo], [Piano intro], [Instrumental].
The model guesses stress on its own
There is no reliable way to set the stress: the model sometimes reads Russian words with the wrong stress, and no trick is guaranteed. Try them one at a time and listen: shorten the line around the word, mark the stressed vowel with a capital («ворОчаются») or spell only the problem word the way it sounds («сонце» instead of «солнце»). It looks odd on paper but sometimes helps by ear.
Artist names are rejected
«In the style of Adele» will be refused. Describe the manner, not the author: «male vocal, minimal arrangement, guitar and voice».
Write repeats out instead of marking them
Marks like «(x2)» are understood unreliably. If a line or a chorus must repeat, write the repeat out in full, as separate lines or a section.
11
What to do with a finished track
Generating is not the end of the work. A finished track can be reworked, and that is often faster and cheaper than driving a new song from scratch.
Extend
If the song cuts off or you want one more verse, extend it instead of creating a new one. Start the continuation with a section tag, [Bridge] or [Chorus], so the model knows what comes next.
Replace a single part
If only the chorus or the instrumental break failed, that is what you replace — there is no need to regenerate the whole track. It is both faster and cheaper than a new song.
Cover and remix
A cover keeps the melody and lyrics but changes the arrangement; a remix reworks the sound more heavily. The source can be one of our tracks or audio you upload.
Vocals and backing track apart
A track can be split into stems: voice separately, music separately. Handy for a karaoke backing track or clean vocals for a video.
Your own voice
You can record or upload a 10 to 30 second sample and sing in your own voice. The profile is saved and applied to later generations.
Helpers before generating
Boost Style expands a short description into a full one, and the AI lyrics button writes the words. They cost 2 ₽ each and save a whole 24 ₽ generation.
Tags are short marks in square brackets inside the lyrics. They are not sung: they tell the model where each part starts and how to sing it. Put each tag on its own line before what it describes and use the standard English names, as in the examples.
Structure: the most important
[Intro] — introduction, [Verse] — verse, [Pre-Chorus] — build before the chorus, [Chorus] — chorus, [Bridge] — a change of mood, [Outro] — ending. Without such tags the model picks a structure from the line breaks and may get it wrong. These are the most reliable and most needed tags.
Delivery
[Whisper], [Spoken Word] — speech without singing, [Rap], [Harmony] — extra voices, [Female Vocal] and [Male Vocal] — a female or male voice. Put a tag where the delivery really changes. It is a hint, not a command: the result is not guaranteed.
Instrumental breaks
[Instrumental Break], [Guitar Solo], [Piano Solo] — a break where the tag is. Describe the overall sound and the set of instruments in the Music Style field, and keep only the mark where a break is needed in the lyrics.
Backing vocals and short cues
Round brackets hold backing vocals and echoes: «I will return (return)», (hey!). Short delivery cues such as (softly) or (whispered) are also used; keep them to one to three words and do not overdo it: ten cues in a song already clutter the text, and full sentences in brackets are best avoided.
The ending
An [End] tag at the end of a song is often used so it finishes without a trailing tail. There is no guarantee: if the song still drags on, set the duration with the slider in the advanced settings.
Tags with a note for one section
Some guides describe a technique: after the tag name, add a short note after a colon that applies to that section only, for example [Verse: whispered vocals, acoustic guitar only] or [Chorus: full band, powerful vocals]. The overall sound stays in the Music Style field. This is the experience of guide authors, not an official feature: the result is not guaranteed, and a long note may be sung aloud.
A markup example: short tags, clean lyrics
[Intro]
[Verse 1]
The night city breathes in neon light,
I walk the empty street.
[Pre-Chorus]
One more step and it all gets clear.
[Chorus]
I will return (return),
When dawn comes up over the river.
[Verse 2]
The windows go dark one by one,
But a light stays on in me.
[Bridge]
(softly) Only you and me,
No one else is here.
[Chorus]
I will return (return),
When dawn comes up over the river.
[Outro]
[End]
Better avoided: long cues before every line
A long tag describing the voice and sound before every line, for example «[Girl 2 - Ethereal dream-pop voice, soft and breathy, cosmic reverb]». The model may sing such instructions aloud, and a crowded text is performed worse than a short, clear one. Describe voices and instruments overall in the Music Style field and keep only short [Verse], [Chorus] and the lines themselves in the lyrics.
13
What to do when the result is off
What you can try when the result is off. Change one thing and generate again so you can see what helped. If a generation fails, the money is returned automatically.
Words are unclear, odd sounds appear
The model may sing instructions written into the lyrics, and a crowded text is performed worse than a short, clear one. Try this: move the overall description of voices and instruments to Music Style, keep short tags in the lyrics and shorten the text.
A tag was sung aloud
Check that the tag is in square brackets on its own line and has a standard name ([Verse], [Chorus]). If it is still sung, shorten it, remove it or move the instruction to the Music Style field.
Wrong stress
There is no guaranteed fix, and results vary. You can try shortening the line around the word, marking the stressed vowel with a capital («ворОчаются») or spelling only the problem word the way it sounds («сонце» instead of «солнце»). More in «What the model does not understand».
The style seems ignored
Put the main genre first, remove descriptors that contradict each other (with ten or more they start to compete) and raise «Style weight» in the advanced settings: the higher it is, the closer the model stays to your description.
The song runs too long or does not end
Set the duration with the slider in the advanced settings and put [End] at the end of the lyrics: this tag is often recommended for a clean ending, but there is no guarantee.
Wrong voice, or you need a duet
Pick «Vocalist gender» in the settings and write «female vocal» or «male vocal» in the style. There is no duet button: you can split the verses with [Female Vocal] and [Male Vocal] tags, but the result is not guaranteed.
The generation was rejected over names or lyrics
Names of real artists in the style and lines close to well-known songs are not accepted. Replace the name with a description of the manner («male vocal, guitar and voice») and rewrite the lines in your own words. If an uploaded file is reported as too similar to an existing work, use another file or your own recording.
The voice profile disappeared
Voice profiles are kept only for a short time and may stop working; they cannot be extended. Keep your recording on your phone or computer: to create the voice again, upload it and confirm with the verification phrase, no need to record from scratch.
14
Rights and using your tracks
In short, as written in our FAQ, without promises we cannot make.
What you can do
You may use generated results in personal and commercial projects within the limits allowed by law and by the rules of the technologies we use.
What you are responsible for
Before publishing, selling, licensing or monetizing, check for yourself that the track does not infringe third-party rights or platform rules. If someone else's lyrics, melodies, voices or images were used as a basis, you are responsible for the lawful use and publication.
What we do not promise
We do not claim copyright in generated results, but we also do not guarantee that there will be no claims from rights holders or platforms. The full terms are in the FAQ.
It is rarely perfect on the first try. First find the direction, then refine the style, lyrics, and mood. This workflow usually gives better results than trying to fit everything into one overloaded prompt.
