This is about using the app. Installing it, and everything you need before it runs, is in the README.
The app writes songs with YuE2, a model that works in two steps: first it writes a score plan — the melody, the chords and the sections, as text — and then it renders that plan into audio. Almost everything here follows from that split. A plan costs seconds and can be read, edited and thrown away; a render costs minutes. So the app lets you look at the plan first.
- Press + Song at the top of the left panel. The editor window opens.
- Type a style. A sentence, not a tag list: "warm indie rock, expressive female lead, jangly guitars, 96 BPM". This is the single biggest influence on what comes out.
- Add lyrics, or press Write lyrics and let Gemma draft them from a description.
- Press Write score plan. Nothing is rendered yet: the window turns to its Score page.
- Read the plan when it lands. If the melody is wrong, Write a new plan rerolls it for the cost of a few seconds.
- Press Render this score. The window closes. A take with no audio yet is filled in; one that already has audio is rendered as a new take beside it, so a render never overwrites one you may want to keep.
Tick render as soon as the plan is ready to run both steps without stopping in between.
The take appears in the library on the right, and plays in the bar at the bottom.
The left panel shows the take you last clicked, read-only: its style, its sound settings, the first lines of its words, and its score as a key, a tempo, a length and each section's chords. + Song, + Cover and + Instrumental at the top start something new; the one for the take's own kind stands out. The chevron beside the panel folds it away, so the takes have the width, and brings it back. Single keys act on the take the panel shows (M for the mastering rack, E for the editor and more): see Hot keys.
Making and changing a take happens in the editor window. Its main button on the panel says what fits: Edit and render again, or Review the plan and render when a plan is waiting. A card's Score and Again buttons open it too, and so does double-clicking a card. It has three columns (the song, the words and the sound), everything in view, with the score on a tab of its own.
Closing the window, with Close, the cross or Escape, keeps what you typed.
The button after Close, in the top right corner, makes the window fill the whole browser window, and a second click puts it back to its usual size. The window opens the way you last left it. Filling the window, the words column takes its full height: the lyrics box (once there are words in it) and the structure grow into the room, and the player stays at the foot of the column.
The window's top row has the take's own small buttons from its card, larger: Sing again, Variations, Try more, Inside, Take out, Normalise and Move, each where the card would show it. Normalise is green while the take is normalised, and a click then puts it back to the level it was rendered at; Inside is lit while the take holds others, and closes the window to show them under the take.
The Advanced button in the editor's header opens settings for this one take. Most songs never need them. They are saved with the take, so Edit and render again finds them as you left them. The button lights up when any setting differs from its default, and Reset to defaults puts them all back.
Some of these act on the plan, and some on the render and the finished file:
- Diffusion steps (16–64, default 32): how many steps the audio decoder takes. More steps take longer to render.
- Avoid: words to steer away from, such as "drums" or "rap vocals". They are added to the style as "avoid: …". It is a hint to the planner, not a guarantee: the renderer has no negative prompt.
- Key lock: the plan is moved to this key once the planner has written it. The planner ignores a key named in the style, so this is done to the score itself: every note and chord shifts by the same interval and the key signature is rewritten. A minor plan stays minor, and only the tonic moves. A plan the app cannot read with certainty, such as one in a modal key or with a key change part way, is left in the planner's key and the log says so.
- Tempo lock (BPM): the plan is set to this tempo, replacing any tempo in the style or in the planner's own score, and the render follows it.
- Score token cap (default 8192): the most the planner may write. A lower cap stops a short song's plan running on.
- Chord hold limit (bars, default 8): the most bars the planner stays on one chord root before it must change.
- Outside harmony bonus (0–10, default 0): how strongly the planner is pushed toward borrowed chords and roots outside the home key. These two act on the planner's choices, so they show as a difference in the chords of the plan rather than as a clear difference by ear. Compare the Score window.
- Voices follow the cues (a tick under the lyrics box, shown once the lyrics have a voice cue such as
[male voice]; on for a new take): the melody of each cued line is moved by octaves to where the voice its cue names is sung. See The words. - Follow this structure exactly (a tick under the lyrics box, not in Advanced; songs; off by default): the plan writes exactly the sections the lyrics name, in order, and no others. Left off, the planner shapes the song itself and often adds an intro, an outro or an interlude, or leaves a section out. A plan can still end before the list does.
- Sections open differently (Follow Harmony, Always, Never): whether a section may open on the same chords as the one before it. Follow Harmony turns it on from Colourful up and leaves it off below; Always and Never decide for this take at any step, so two plans can be compared with and without it.
- Target loudness: the level a take is normalised to. It applies only when normalising is on.
- Outro fade (seconds, default 3): the length of the fade when a take is cut off at the length cap.
YuE2 reads this as a description of a recording, so describe a recording. Language, genre, voice,
instruments, mood, tempo and production, roughly in that order, all in one sentence. Bare tag lists
work less well, and section markers such as [Verse] belong in the lyrics, not here.
The Vocal chips below write into the style for you: female, male or duet, and a character such as breathy or raspy. They are a shortcut for typing, and you can edit the result by hand.
The structure. Above the lyrics box, while it holds no words, a song has a structure to build, as an instrumental has: a list of sections you can reorder, remove and add to (intro, verse, pre-chorus, chorus, bridge, interlude, outro). Start from fills it with a common shape, and each shape's tooltip says what it suits. An interlude is an instrumental passage: the writer puts its tag on a line of its own, and the planner plays it with the instrument voice. Put sections in the lyrics box drops the tags in as empty sections to write under by hand. Once the box has words, the lyrics are what is sung and the structure steps aside.
Write lyrics, beside the lyrics box, asks what the song is about and writes words for the sections you built. Lines in a verse or chorus sets how long each is: more lines in a section make a longer song, because the planner writes about as much music as there are words. The Style above sets the mood. Gemma writes the draft on the same engine, and it lands in the box with a title if you had not given one. The window closes as soon as the draft starts, and the progress shows beside the lyrics box. The lyrics editor's buttons add any of these sections too, along with voice cues (+ Male, + Female, + Both).
How the lines sound sets the rhythm and rhyme the writer aims for. Natural (the default) mixes short and long lines with some rhymes. Regular rhyme is the traditional shape of even lines that rhyme in pairs, which is what the writer produced before this choice existed. Spoken rhythm is conversational, like speech or rap; Loose rhyme rhymes only some lines; Short and punchy and Long and flowing set the length of the lines.
It is a first draft. The lines scan and rhyme, but a model reaches for familiar images, and nothing checks whether a line is already someone else's. Read it and make it yours before you plan.
+ Song has a lot of choices now. None of them is required: a style and some lyrics are enough, as in Your first song. This chapter goes through all of them in the order the editor shows them, its three columns (the song, the words and the sound), and then what to do after the plan. It ends with a table for the common wishes, such as a longer song or a chorus that sounds different from the verse.
- Title. Optional. Left blank, the first line of the lyrics is used.
- Style. The Choose a style… list holds 200 starting points, from afrobeats to zydeco, in alphabetical order. Picking one fills the Style box with a full description, which you can then edit. The chips under it are ten quick starts for a song: warm indie rock, soulful jazz-pop, synthwave, acoustic ballad, heavy rock, modern pop, country, hip hop, smooth R&B and dance pop. See What to put in the Style box.
- Saving a style of your own. Save, beside the list, keeps what is in the Style box under a name you give it. The list's first entry, Saved styles, opens onto the ones you kept; the bin beside one deletes it, and the list's search finds them along with the 200. Saving under a name you have used replaces that style. A style LoRA's trigger word is left out of what is kept, and put back for whichever LoRA is chosen when you use the style.
- A tempo in the style is only a hint. The planner often picks its own. To fix the tempo, use Tempo lock in Advanced.
- Vocal. The chips write a voice (female, male, duet and a character) into the style for you.
- Style LoRA. Under the sound column. See Style LoRAs.
The structure. Above the lyrics box, while it holds no words, there is a list of sections: the order the
song will go in. Each row can be changed to another section, moved up or down (or dragged), or removed, and the
chips under the list add a section. The sections are intro, verse, pre-chorus, chorus, bridge, interlude and outro.
An interlude is an instrumental passage: it has no words, and the planner plays it with the instrument voice.
(solo is not a section the planner knows; interlude is the way to ask for one.)
Start from… fills the list with a common shape. Hover an option to see what it suits:
| Shape | Sections | Suits |
|---|---|---|
| Verse, chorus, bridge | verse, chorus, verse, chorus, bridge, chorus | the classic song: pop, rock, country |
| Short | verse, chorus, verse, chorus | folk, punk, early rock and roll |
| Intro and outro | intro, verse, chorus, verse, chorus, outro | ballads, soft rock, singer-songwriters |
| Story | verse, verse, chorus, verse, chorus, outro | story songs, folk, country |
| Pop with pre-chorus | verse, pre-chorus, chorus, verse, pre-chorus, chorus, bridge, chorus | modern pop, dance-pop, pop rock |
| Hook first | chorus, verse, chorus, verse, chorus, bridge, chorus | streaming-era pop, indie pop |
| Short and simple | verse, chorus, verse, chorus, chorus | punk, garage rock, rockabilly |
| Hip-hop | intro, verse, chorus, verse, chorus, verse, outro | rap, hip-hop, trap, R&B |
| Double bridge | verse, chorus, verse, chorus, bridge, bridge, chorus, chorus | power ballads, progressive and symphonic rock |
| Interlude | verse, chorus, interlude, verse, chorus, bridge, chorus | rock, pop rock, indie |
| Every part | intro, verse, pre-chorus, chorus, interlude, verse, pre-chorus, chorus, outro | pop, dance-pop |
How many sections. The planner usually writes 8 to 10 sections whatever the lyrics ask for, so a list longer than 10 gets condensed and the words of the sections it drops are not sung. The builder says so when the list is that long. A song does not get longer by adding sections; each one just gets shorter. See Making the song longer.
Put sections in the lyrics box drops the sections in as empty tags, ready to write under by hand. Once the box has words, the lyrics are what is sung, and the builder steps aside.
Write lyrics. Describe what the song is about and a model writes words for the sections you built. It asks for three things:
- Lines in a verse or chorus (4, 6 or 8; 6 is the default). More lines in a section make a longer song, because the planner writes about as much music as there are words. An intro, bridge or outro stays at 2 to 4 lines, and an interlude has none.
- How the lines sound. The list shows the syllables each line aims for. The old writer produced the same eight-syllable rhymed lines nearly every time, so this changes that:
| Choice | Aims for | What you get |
|---|---|---|
| Natural (the default) | 3-13 syllables | A mix of short and long lines, some rhymes, no steady beat |
| Regular rhyme | 6-10 syllables | Even lines that rhyme in pairs: the traditional shape, and what the writer used to do |
| Spoken rhythm | 5-14 syllables | Conversational and rhythmic, like speech or rap, with internal rhymes |
| Loose rhyme | 4-13 syllables | Some rhymes, some slant rhymes, some none, and lines of different lengths |
| Short and punchy | 3-6 syllables | Short lines and repeated phrases |
| Long and flowing | 10-14 syllables | Long lines that run on like a thought |
The last choice is remembered. These are aims, not guarantees: a language model follows them closely for the extremes (short and punchy, long and flowing) and more loosely for the subtle ones.
- The brief itself, in your own words, and the Style above, which sets the mood. The brief is kept with the take, so opening a take and pressing Write lyrics shows what it was asked to be about, ready to change and write again.
The draft lands in the lyrics box, and the window closes as soon as it starts. It is a first draft: read it and make it yours.
Follow this structure exactly (the tick under the lyrics box, off by default). The planner treats the section tags in the lyrics as guidance, and left alone it writes its own shape: it often adds an intro and an outro, swaps a bridge for an interlude and a verse, or condenses a long list. With this on, the plan writes exactly the sections the lyrics name, in that order, and no others. It cannot add an intro or an outro that you did not list. The planner still decides what the music is and when each section starts. It applies to songs only.
Who sings each line: voice cues. A line on its own in square brackets that names a voice says who sings the lines under it, until the next cue or the next section:
[Verse 1]
[male voice]
The ferry leaves at half past nine
I watch the harbour lights go by
[female voice]
I wait beside the ticket line
[Chorus]
[both voices]
Carry me over, carry me home
Write a cue as [male voice], [female voice] or [both voices], each on a line of its own in square brackets.
The expanded lyrics editor has + Male, + Female and + Both buttons in its header to insert them at the cursor.
These are the forms the feature was tried with, along with vocal for voice ([female vocal]) and
[male and female voices] for both. You can add words of your own after a comma, as in
[female voice, soft and close]; they do not change who sings. A section with no cue is left to the planner.
Other spellings are read as well, so lyrics written elsewhere are not ignored: [man], [woman], [duet], a cue
inside a section's tag such as [Verse 2 - female vocal], and a cue in parentheses, (female voice). Square
brackets are the safer habit: words in parentheses are usually backing vocals, and may be sung.
A cue alone does not decide the voice. The singer comes mostly from how high the melody is written: the same line an octave up is sung by a woman, and the planner likes to write a verse in two octaves, changing every four lines or so, whoever was cued. Voices follow the cues (a tick under the lyrics box, shown once the lyrics have a cue; on for a new take) keeps the cues: when the plan is written, each cued stretch of the melody is moved by whole octaves to where that voice sings, low for a man and high for a woman. Nothing else in the plan changes. It is done again before each render, so it follows a change to the words or to the score. In the Score page, Voices to cues does the same once to the score in the box, for a take made without the tick; save the score and render it.
Three things to know:
- Who sings a line is still partly chance. With the melody in the right place most lines go to the voice cued, but a render can give one to the other singer all the same. Render it again with another seed, or use Try more.
- Both puts the line where a woman sings and leaves the second voice to the render, which often brings one in for a chorus. It asks for a duet rather than making one: how much of the second voice you hear varies from render to render. There is no way yet to write two singers on one line without the words suffering.
- A take made before this tick existed keeps its score as it was; tick it, or use Voices to cues, to fit it.
Writing by hand. The Expand button opens a full-size editor with buttons for intro, verse, pre-chorus,
chorus, bridge, interlude, outro and instrumental. A section is a line such as [Verse] followed by its lines; a
tag with no lines under it, such as [Interlude], is an instrumental passage.
-
Harmony (the slider): how adventurous the chords are. A new song starts at Varied, with the interpretation set to Wide: in our own testing that pair gave the most musically interesting renditions, so it is a good place to start. Move either to taste. Existing takes keep what they were made with. The planner left alone tends to write one four-chord loop and use it for every section, and each step pushes it a little further from that. The steps build on each other:
Step What it does to the chords Reach for it when Familiar Nothing: YuE2's own chords, often one loop for the whole song You like the planner's first idea, or want the plainest backing Varied Discourages repeating the exact chords it has just used, so it recolours them (an F, then an Fmaj7) but stays in the key The loop feels too samey, and you want the song to stay conventional Colourful The same, harder: verse and chorus part ways, with richer chords. A section also may not open on the same chords as the one before it, so a chorus no longer starts the way its verse did You want the chorus to feel like a lift, not a repeat. A good first step up Adventurous Discourages recently used roots, not just spellings, so the harmony has to travel, and it borrows chords from outside the key You want movement and surprise, and will listen for a chord that sounds odd Outside Adventurous, with a pull towards chords outside the key Experiments. Expect to reject some plans From Colourful up the chord choices come out more distinct from section to section: more different chords in a song, more four-bar patterns that are not repeats, and less of the chorus copying the verse's opening. The plan is still a normal song, so it is worth trying a step or two higher before reaching for a wilder Plan variety. Under Advanced, Sections open differently can be set to Always or Never to try a plan with and without the chorus rule; the default, Follow Harmony, is on from Colourful up. A style such as trap or ambient is often loop-based by nature, so expect a smaller change there than in pop or rock. See Harmony for the measurements.
-
Plan variety. How freely the planner writes everything else. See Plan variety.
-
Interpretation. How the notes are performed when rendered. See Interpretation.
-
Length cap. The latest the song may stop. See Length.
-
Mode. Full keeps the chords of the plan; melody leaves the accompaniment free.
-
Seed, and keep this seed. See Seed, and reproducing a take.
-
Production polish and Normalise volume. See Production polish.
-
Render as soon as the plan is ready runs the plan and the render without stopping in between. Leave it off to read the plan first.
-
Advanced (the button in the header): diffusion steps, Avoid, Key lock, Tempo lock, the longest score, chord hold limit, outside harmony bonus, Sections open differently, target loudness and the outro fade. See Advanced settings.
-
Mastering (the button to its left, when the take has sound) opens the mastering rack over the editor for the take in it. The rack is, in order, a 1073-style EQ, a parametric equalizer, a compressor, a stereo imager and a limiter. On the parametric equalizer's graph each of the seven bands is a dot: drag it across for its frequency and up or down for its gain, turn the wheel over it for its width, double-click it to put it back. The strip under the graph sets the chosen band's kind (low cut, low shelf, bell, notch, high shelf, high cut) and switches it on or off, Flat puts every band back to 0 dB, and the button in the graph's corner changes how many decibels the graph shows. The button is green while the rack changes how the take sounds. With the rack open, space plays and pauses the take and the arrow keys move through it and set the volume, as on the main page. Escape closes the rack, and again closes the editor.
The length of a song is set by its plan, and the render follows the plan's length. These are the things that move it:
| To get… | Do this |
|---|---|
| A longer song | Give each section more lines (Lines in a verse or chorus: 6 or 8) |
| A slower, longer song | Set a Tempo lock in Advanced. The same plan at a lower BPM runs longer |
| A longer song by adding sections | This does not work: a plan has about 8 to 10 sections, so more just get shorter or are condensed |
| More than the plan allows | Write a new plan; lengths vary a lot between runs on the same words |
| A song that stops at the end of its words | Raise the length cap only if it is being cut short; the cap is a limit, not a target |
- Render this score makes a new take whenever the open take already has audio, so a render never overwrites one you may want to compare. A take with no audio yet is filled in.
- Keep this tune (after changing a take's words): sings the take's score with the new words, as a new take. The tune cannot gain sections or notes, so if your words have more sections than the tune does, or far more syllables than it has notes (as with Long and flowing lines on a tune made for short ones), the tick clears itself and a new plan is written, with a line saying why. See Same tune, new words.
- Filter the library. The Song, Cover and Instrumental toggles beside the filters show only those kinds in the open space; pick more than one, or none for every kind.
The plan is written in ABC notation: music as plain text, letters instead of dots on a stave. Chris Walshaw devised it in the early 1990s for sending folk tunes by email, and it stuck.
You do not need to read it. The app renders the same plan as a chord chart and as real staff notation below the editor, and Expand opens all of it full size. But the text is what YuE2 wrote and what it renders from, so it is worth being able to find your way around.
A plan begins with a header, one letter and a colon per line:
X:1
M:4/4
L:1/32
Q:1/4=120
V: Vocal clef=treble name="Vocal Melody"
V: Ins clef=treble name="Ins Melody"
K:D#m
| Line | Means |
|---|---|
X:1 |
the tune's number. Every ABC file starts with one |
M:4/4 |
the metre — four beats in a bar |
L:1/32 |
the unit length: a bare letter lasts a thirty-second note |
Q:1/4=120 |
the tempo — a quarter note at 120 beats per minute |
V: |
declares a voice. YuE2 writes two, a vocal and an instrumental line |
K:D#m |
the key, D sharp minor. K: always comes last in the header |
Then the music itself:
% intro
V: Vocal
z24z4"D#m"z4|"D#m"z32|
V: Ins
Z|d16a16-|a16g8a8|
- Notes are letters.
d16is a D lasting 16 units. Lowercase sits an octave above uppercase, and'or,shift it further. -ties a note into the next one, so they sound as one.zis a rest, with a length like a note. A capitalZrests for a whole bar.- Chords live in double quotes, like
"D#m", and sit in front of the note they start on. These are what find and replace edits. |is a bar line, and%starts a comment — which is how the sections are marked:% intro,% verse,% chorus.
So that fragment says: D sharp minor, four four, 120 beats per minute, and through the intro the vocal rests while carrying the chord, while the instrument plays a tied D and A, then G and A.
Two practical consequences. Editing chords is safe even if the notes look opaque — they are only the parts in quotes, and everything else can be left alone. And if you edit the plan, the render uses what you edited, so repairing one is usually faster than rerolling until a good one appears.
The Notation tab has a player above the staves. Press Play to hear the plan as the box has it, edits and all, so a melody you are unsure of can be judged by ear before a render is spent on it. The notes it is playing are picked out as it goes, and its tempo control changes the preview only, not the score. Playing it stops whatever the player bar had, since two things at once is no use to anyone, and closing the window — or moving to Chart or Lyrics — stops it too, so nothing is left playing with no control in reach.
Instruments plays each voice with an instrument the Style names, rather than a piano throughout: a style that says electric guitar, drums is heard with a guitar, its chords played by another, and a bass under them; a sung line is stood in for by a voice, and an instrumental's melody is not. A style that names drums gets a drum part too, written against the score's own metre — a backbeat for rock, four to the floor for dance, a ride for jazz, something sparser for a ballad — and it plays all the way through, since the score has no sections for it to sit out. The line beside the player says what it chose, and unticking it goes back to the piano you would otherwise hear. This is a reading of the style, not of the render: YuE2 decides the real performance and adds parts the score never had, so treat it as the plan played in the right colours rather than as what the take will sound like.
Download MIDI saves the score for a DAW. Each part is a track of its own, on its own channel, with the instrument the preview used set on that channel: the vocal line, the instrument line and, unless you untick Chords, the chord symbols as a chord part with its bass beside it (the only place the harmony is written as notes), and drums on channel 10 when the style asks for them. The instruments are General MIDI program numbers, so your DAW plays them with its own sounds, and the tempo and metre are in the file. A plan that changes its metre part way is written in the metre it starts in — only that one is recorded — while the notes keep their own lengths. Clicking a note in the staves puts the cursor on the ABC that wrote it.
A score with a note above what these instruments can play, or below it, is played shifted by whole octaves until it fits, and the line beside the player says how far; the staves, the MIDI file and the score itself are unchanged. The instruments are not all the same size — a bass stops well below a piano — which is why the line is there when a score is shifted. A score with no notes in it yet has nothing to play and says so.
The player needs its note samples, which are not part of the app: each instrument is about 7 MB, fetched from its upstream home the first time a preview plays with it and kept in the library, so it works offline afterwards. Play fetches whatever the current score needs and then starts, and Get the sounds fetches the same without playing, for anyone who would rather have it ready first.
Click Piano Roll in the score window to open the visual MIDI editor. Rather than editing raw ABC letters, you can see and shape the melody directly on an interactive pitch and time grid.
The editor displays both score voices simultaneously — Vocal (the lead singing melody in sky blue) and Ins (the instrumental accompaniment in warm amber). Both voices are always fully active, visible, selectable, and editable. The Draw buttons at the top left choose which voice new notes will be assigned to when you click to add notes.
- Add a note: Click any empty cell on the grid to create a note at that pitch and time. The note auditions immediately so you can hear its tone.
- Select notes: Click a note to select it. Hold Shift or Ctrl/Cmd while clicking to select multiple notes across both vocal and instrumental parts. Click and drag a marquee selection box over empty grid space to select any phrase or passage across all voices. Press Ctrl+A (Cmd+A on macOS) or click Select All to select all notes across the entire score. To select all notes from the playhead cursor to the end of the song, press Ctrl+Shift+A or Alt+A (or click From Cursor ▶ in the toolbar).
- Move notes: Drag any selected note to shift its timing earlier or later, or drag vertically to transpose its pitch. When multiple notes are selected — whether vocal, instrumental, or a combination of both — they all move together in lockstep, preserving their relative timing, melodies, and lyrics.
- Resize and duration: Drag the right-hand edge of a note to lengthen or shorten its duration. The Snap selector (1/16, 1/8, or 1/4 note) constrains movement and resizing to clean musical divisions.
- Delete notes: Double-click any note to delete it, or select one or more notes and press Delete or Backspace.
- Undo and redo: Every note edit, move, resize, addition, and deletion is recorded in the history stack. Press Ctrl+Z (Cmd+Z) to undo, and Ctrl+Shift+Z or Ctrl+Y to redo.
- Audition keys: Click any key on the left-hand piano keyboard to hear its pitch.
- Maximize: Click Maximize in the top toolbar to expand the piano roll to the full width of your browser window.
Along the top ruler, the timeline displays measure numbers and section boundaries:
- Rename sections: Click any section marker badge (
intro,verse,chorus, etc.) to edit its name or reassign it. - Edit chords: Directly below the bar ruler, the chord track displays the song's harmonic progression. Click an existing chord badge to rename it (for example, change
CtoAm7orG/B), or click an empty bar slot to place a new chord. - Harmonize empty bars (Fill Gaps): Click Fill Gaps in the toolbar to populate empty bars in the instrumental accompaniment track with musical notes generated from the chord progression.
- Remove empty bars (Compact Gaps): Click Compact Gaps to remove silent empty bars across both voices, shifting subsequent notes and chords left to eliminate unwanted gaps in the arrangement.
At the bottom of the piano roll, a dedicated lyrics track aligns each syllable with the vocal melody:
- Phrase-aware matching: Click Match Lyrics to distribute lyrics across the vocal melody. The alignment detects natural musical breath pauses and rest gaps between melodic phrases, assigning each lyric line to its intended phrase without spilling words across rests into subsequent measures. Multi-syllable words split cleanly across notes.
- Editing syllables: Click any lyric tag (or select a vocal note and press L) to edit its text. If you enter space- or hyphen-separated syllables (for instance,
Hel- lo world), they automatically flow across consecutive vocal notes. - Hover highlighting: Hovering over a lyric tag highlights the corresponding note on the piano roll grid, and hovering a vocal note highlights its lyric tag below.
- Karaoke dancing ball: During playback, an animated dancing ball arcs across each note as it is sung, and the active syllable and grid note illuminate with a real-time glow.
Press Space or click ▶ Play to start playback from the playhead cursor. The playhead line sweeps across the grid, updating the current bar, beat, and elapsed time counter.
- Audible click track (metronome): Playback includes an audible click track by default, with an accented woodblock click on beat 1 of each bar and softer clicks on inner beats. Press C or M (or click Click in the transport bar) to toggle the metronome on or off at any moment.
- Harmonic chord accompaniment: Playback synthesizes warm polyphonic chord accompaniment pads directly from the score's chord progression. Even when the vocal or instrumental tracks rest, the harmonic space remains musical and warm rather than dropping out into dead silence. Press H (or click Chords in the transport bar) to toggle chord accompaniment on or off.
- Stepping measures: Press Left Arrow and Right Arrow to step backward and forward bar by bar, or press Home to jump straight back to the beginning.
| Key | Action |
|---|---|
| Space | Play / Pause playback |
| Left Arrow | Step back one bar |
| Right Arrow | Step forward one bar |
| Home | Return playhead to start (bar 1) |
| C or M | Toggle audible click track (metronome) on / off |
| H | Toggle chord harmony accompaniment on / off |
| Delete / Backspace | Delete selected note(s) |
| Ctrl+Z / Cmd+Z | Undo last edit |
| Ctrl+Shift+Z / Ctrl+Y | Redo edit |
| Ctrl+A / Cmd+A | Select all notes across both voices |
| Ctrl+Shift+A / Alt+A | Select all notes from playhead cursor to end across both voices |
| + or = | Zoom in timeline horizontal scale |
| - or _ | Zoom out timeline horizontal scale |
| F or 0 | Scroll to notes (Fit view) |
| L | Edit lyric for selected vocal note |
| Escape | Clear note selection, or close score window |
YuE2 left alone tends to write one four-chord loop and stay there. The Harmony slider pushes it away from chords it has just used, without breaking the song's structure.
| Step | What you get |
|---|---|
| Familiar | YuE2's own chords. Often one loop for the whole song |
| Varied | Avoids repeating the same chords. Stays in the key |
| Colourful | Verse and chorus get different progressions, with richer chords. A section may not open the way the one before it did |
| Adventurous | Keeps the harmony moving, and borrows chords from outside the key |
| Outside | Adventurous, and reaches further outside the key |
Write a new plan uses the slider's current position, so you can reroll the same words with more adventurous chords and compare.
This was measured rather than guessed — one set of lyrics, two styles, three seeds each:
| Familiar | Varied | Colourful | Adventurous | Outside | |
|---|---|---|---|---|---|
| Different chords in a song | 4.7 | 6.3 | 8.7 | 9.0 | 10.3 |
| Bars using a chord from outside the key | 0% | 0% | 0% | 13% | 24% |
| Four-bar patterns that are not repeats | 29% | 38% | 50% | 63% | 64% |
Every plan at every step kept its sections and valid chords. Asking for adventurous harmony in the style text instead — "jazz harmony", "borrowed chords" — had no measurable effect at all.
Under Advanced, Plan variety sets how freely the planner writes. Where Harmony acts on the chords, this acts on everything: melody, structure and length.
- Calm: the steadiest. Few chords, and a melody that stays close to home.
- Normal: the default.
- Lively: a few more chords and turns.
- Bold: richer harmony, about three times normal's chord vocabulary, while the tune stays recognisably the same kind of tune.
- Quirky: restless. Chords keep changing, and the melody leaps between registers from one section to the next. For odd, off-the-wall songs.
- Wild: the most restless of all, and it may change key partway through.
The score fixes the notes. The interpretation sets how they are performed.
| Interpretation | What you hear |
|---|---|
| Standard | YuE2's usual reading |
| Tight | more controlled and polished |
| Loose | rougher and more spontaneous |
| Settled | free to repeat a figure and sit in a groove |
| Restless | keeps the parts moving, avoids repeating itself |
| Wide | reaches for less obvious sounds |
The sparkle button on a card renders the same score and the same seed in the other interpretations. Because only the interpretation changes, what you hear between them is the interpretation — not a different roll of the dice. Each is a take of its own, titled Night drive · Loose and kept inside the take it was made from (see Takes inside a take, under The library). The window has its own length cap, which starts at the editor's and applies only to these takes.
The dice button on a card renders the same score and words again as several new takes, to listen to together. Variations changes only the interpretation and keeps the seed; this changes the roll:
- New seeds: a fresh seed for each, everything else kept.
- Planner strengths: for a take with a style LoRA that has a planner half, the same seed at other Planner strengths (its own is left out). Planner changes the performance itself, not only the voice, so it works like another seed.
The new takes are titled Night drive · try 1 or Night drive · planner 0.60 and are kept inside the take they were made from, whose panel opens to show them as they finish. To hear them against each other, press Select all while the panel is open and then Compare. The window has its own length cap, which starts at the take's own.
Every take records its seed and every setting that shaped it. The card names them, clicking a card shows them in the take panel, and opening it in the editor loads all of it back. Again opens the editor with exactly that, seed included, so a take you liked can be reproduced, and a take you nearly liked can be nudged one setting at a time.
In the editor's menus, hovering an interpretation shows what it does.
Tick keep this seed to keep the same seed across renders; leave it off and each render rolls a new one.
A render uses what the editor shows: its style, style LoRA and strengths, length cap, mode, interpretation and seed, even when rendering a take's own score again.
The planner reads all the lyrics before it writes a note, so the same seed gives the same plan only for the same words. Change one line and the next plan is a new tune: another melody, often another key and tempo, and the voice moves with them.
To keep a tune you like, open the take and change its words. Keep this tune appears under the lyrics, ticked, and the main button becomes Sing with new words. Change anything else you want first: the style, the LoRA and its strengths, the length cap, the interpretation. The take's score is then sung with the new words, as a new take beside the original, which stays as it is. Keep the seed ticked to keep the voice as close as it can be. Untick Keep this tune to write a new plan, and a new tune, for the words instead. New words that keep the old lines' syllable counts fit the melody best. When they do not, because there are more sections or far more syllables than the tune has notes (long lines on a tune made for short ones: the lines come out repeated or dropped), Keep this tune clears itself with a line saying why, and a new plan is written for the words.
Production polish applies Mothersuperior's Realaudio decoder LoRA. Stock YuE2 often sounds boxy in the mid-range; this separates instruments and vocals more cleanly. On by default.
The length cap is a firm limit, not a target. YuE2 decides when a song ends, and usually ends by itself; the cap stops one that will not. It starts at 360 seconds, or at what Length cap for a new take in Settings says. It then makes room for the music, so a long song is not cut short:
- A song or instrumental from a prompt has its cap raised when its plan is written, if the plan is longer: to the plan's length with about ten seconds to spare. It is never brought down.
- With a recording chosen, for a cover or an instrumental, the cap follows the recording's score, rounded up the same way.
A cap you type yourself stays as you set it, and the Score page then says when a plan is longer than it. A plan much longer than its cap (more than a third over, and over eight minutes) is not taken for a song at all, and is written again. Now and then a render doesn't stop at the end of its score and plays on until the cap; such a take is faded out over its last few seconds rather than cut off, and its card says it ran to the cap. Rarer still, a render ends before the last section of its score begins. The app then renders it once more with a new seed before calling it finished, and if that one stops short too, the card says it stopped before the last section: Render again for another try.
- Drop in an audio file, up to 300 MB. It is stored once and hashed, so the same file is never held twice.
- Press Transcribe. SheetSage2 writes the melody and the chords into the score box. This is cached per recording, so covering the same song again skips it.
- Fix anything it misheard.
- Add lyrics, choose a style, press Create cover.
A cover follows the original's melody and chords while the style decides everything else, which is what makes it a cover rather than a copy.
Note
Experimental: Using MIDI files is designed for generating different, often off-the-wall takes of original tunes.
In addition to recorded audio files, you can drop or upload Standard MIDI files (.mid, .midi) directly into Cover mode.
When a MIDI file is uploaded:
- Automatic score transcription: Yeufonic analyses tracks and channels to identify the lead vocal melody, accompaniment, and harmonic chord progression, transcribing them into ABC notation and loading them directly into the Piano Roll. Synth leads or vocal melodies sequenced in lower registers are automatically transposed to a natural singing octave.
- Embedded lyrics: Any lyric or text events embedded within the MIDI sequence are extracted and aligned to the melody syllables.
- Where the bars fall: a MIDI file has no bar lines of its own, and some start with the first word on the beat before the downbeat. Yeufonic reads the drums and bass to find the real bar line and, when the file starts with such a lead-in, puts the first words in a short bar of their own so the chords and bar markers land where the song's do. A file with no clear drum and bass pattern is left as it is.
- Pasted lyrics: words you paste are fitted to the whole melody at once, a syllable to a note, with the notes left over held on the syllable before. Lines end where the melody rests and begin where a phrase begins, so a section that is longer or shorter than the lyrics' headings no longer pulls the rest out of step. If the words are longer than the melody can carry, the last lines are left out.
- Full audio audition with SoundFonts: The audition player synthesizes the MIDI arrangement into high-fidelity audio using FluidSynth and General MIDI SoundFonts (
.sf2). The output is peak-normalised so that quiet multi-track MIDI recordings audition with full presence and clarity without distortion. - Select SoundFonts: In Settings, you can choose your preferred active
.sf2SoundFont bank (such as Arachno SoundFont 1.0 or JNS-GM 2.0). Additional.sf2files placed indata/models/soundfonts/sf2/are picked up automatically. - Piano Roll refinement: Open the Piano Roll to inspect or modify notes, choose which track acts as the lead vocal melody, harmonize accompaniment gaps with Fill Gaps, or align newly pasted lyrics with Match Lyrics.
A corpus's analysis has already done what a cover needs: the score, the words heard in the song, and its separated vocal. So below your recordings, the list offers From your corpora, folded by corpus, with each song's key and tempo, and a filter for a big one. Pick a song and it becomes one of your recordings at once, with its score and its words, and nothing runs again. The words are laid under the score's sections, as Extract lyrics lays them, unless you checked them in the corpus, when they come as you left them. They go into the Lyrics box, replacing words that came from another recording; words you typed or edited there are replaced only if you say so. The Style box takes the style the song was learned with in the corpus, the same as its chip under a LoRA trained from it.
The file is linked rather than copied where the disk allows, so it takes no more space, and it stays if the corpus is deleted. A few songs are marked score: melody only or score: first 4 min: their full transcription failed during the analysis, which does for training. Press Transcribe for the whole score with its chords.
With a LoRA. This is where corpus songs are most use: a song and a LoRA made from the same work, or from different ones.
- In Cover mode, open the recording list and pick the song under From your corpora. Its score sets the melody and chords, and its words are in the box.
- Choose a Style LoRA. The corpus's own LoRA has its singer cover their own song, closest to the original. Another corpus's LoRA covers it in that other style and voice.
- Set the strengths as for any cover: the recording already sets the melody, so keep Sound near 0.50, and raise it for more of the LoRA's voice and sound.
- The Style box already holds the song's own style, and choosing a LoRA puts its trigger word in front. That keeps the song's genre and mood; another chip, or the filter's tags, moves it somewhere else. The tempo and key come from the score, whatever the style says.
- Press Create cover. Variations then renders it in the other interpretations, and Sing again gives another voice over the same score, as with any take.
Extract lyrics, beside Transcribe, writes down what the recording sings: it separates the vocal, listens to it, and lays the lines under the sections of the score. It is asked for rather than done with every transcription, because it takes a couple of minutes where transcribing a score takes seconds. The vocal is separated, and its words heard, on the GPU when the engine has Demucs and Whisper and on the CPU otherwise (Settings, Use the GPU for stems and lyrics), and the bar says which of the two stages it is on. Whisper on the GPU and on the CPU do not always write the same words for the same vocal, as neither is exact: a hard vocal, such as backing singers over a lead, can differ by a fifth of its words either way. The separated vocal is kept, so extracting the same recording again skips straight to the listening, which is most of the wait saved.
The words are kept with the recording. Press Extract lyrics again and they go straight back in the box, and you're asked whether to extract them again, which is worth doing after changing the method in Settings. Choosing another recording takes its words in the same way, and a recording with none clears the last one's. Words you typed or edited stay unless you agree to replace them.
Expect a good draft rather than a transcript. Measured against the real words of two songs, Whisper got 1.5% of words wrong on one and 25% on the other, where lead and backing vocals sing over each other in the last chorus. It listens to the whole vocal: it used to skip whatever its voice detector took for silence, which on sung vocals was most of the song. Read the draft and fix what it misheard, especially where voices overlap.
Letting an external LLM listen. If an external LLM is your provider, Settings has a choice under Lyrics from a recording. Set it to External LLM and the separated vocal is sent to the model to hear the words, while Whisper still works out when each line is sung. On the same two songs Gemini got 1.5% and 23%, so the two are close; try both on a song Whisper struggles with. It needs a model that accepts audio, such as Gemini. When the model's reply looks like lyrics, its words are used, even where they differ a good deal from Whisper's: a vocal buried in a mix is where Whisper struggles most, and the model does not. If the model refuses the audio, holds back (cutting lines short, or pointing you at a lyrics site, as a model may with a song it recognises) or sends back almost nothing, you get Whisper's version instead. The message when it finishes names which one heard the words. The vocal leaves your machine for this; Whisper keeps it here.
For a corpus song, both versions are kept. When both exist, the song's Review panel shows Words from with each version and its word count, the external model's in use by default. Pick the other and its words go in the box. The first time a version is chosen its words are drafted, which marks their sections with the external model; after that each version keeps its own words, edits and checked tick, so switching between them is instant and makes no call to the model. Whisper's own lines have any word it repeated more than eight times in a row, such as a held "la" it wrote hundreds of times, cut back to eight.
The third mode writes a piece with no vocal. In place of lyrics it takes a structure:
| Structure | What YuE2 gets | Who decides |
|---|---|---|
| Let YuE2 decide | [instrumental] |
YuE2 chooses the sections and their lengths |
| Sections | [intro] [verse] [chorus] … |
you choose the sections, YuE2 their length |
| Timed sections | [intro 0:00-0:15] … |
you choose both |
Add sections with the + chips, reorder them with the arrows, and give each a length when timed. Sent to YuE2 shows exactly what the model receives. The structure is guidance: YuE2 may rename a section, add an interlude, or run past the times you gave, so the length cap is the firm limit.
From a recording at the top of the editor takes one of your recordings, as a cover does. Its transcription is the score, so there is no plan to write, and the render plays its melody and chords with the instrumental LoRA.
The sections come from the score too: one per section of the score, in the names the LoRA knows. An interlude becomes a bridge. A render pairs each section of the structure with one of the score, which is why the builder gives way to them. To rearrange the song, drag a section to a new place or move it with the arrows, add a copy of it after itself, or take it out with the ✕: each change rewrites the score (the Score window shows it), and Restore the original sections puts it back until another score is loaded. While the take plays in the editor, the section it is at is lit, and double-clicking a section jumps to it. The times come from the score's bars and tempo, which the render follows closely for the first part of a take and for one that renders in full. A render that ends well short of its score has played less music than the score holds, so the list can drift out of step with the audio as the song goes on; a line under the list says so and says by how much. A cover shows the same list under its words, and its words are matched to the sections in order, so change them to suit: the list says what they need after you add, remove or rearrange sections. It also shows the total against the length cap, and offers to raise the cap when a longer song would be cut short. Press Create instrumental to render.
An instrumental you wrote from a structure has a score too, and once it is planned its sections are listed in the same way. The structure builder is the brief for a plan not yet written, so it gives way to the list. Rearrange the sections and press Render this score: the render's section tags follow the edited score. A take that already has audio is rendered as a new take beside it, whatever you changed, so the original keeps its own. Write a new plan (under the score) still asks for a different melody from the structure the take was written with.
A song with vocals works too. Its score has the sung melody in the vocal part, which would come out sung, so the tune is given to an instrument: wherever no instrument is playing, the sung notes move to one, and the vocal part keeps only its chords. The editor says so under the recording. It is usually the better choice for a song: without its voice, the transcriber tends to miss the difference between verse and chorus, and hears the whole song as one loop of chords. The LoRA arranges it as an instrumental, so expect parts of its own, such as extra guitar.
Choose No recording in the list to go back to writing a plan.
Occasionally the model puts a voice into an instrumental. This is a model failure, not a setting you got wrong. Before rendering, if the plan puts notes in the vocal voice, a dialog offers a new plan, a new seed, or rendering anyway. A finished take is not checked for singing: listen to it.
Make an instrumental, under a song's or cover's score on the editor's Score page (beside Render this score for a song), makes the same score as an instrumental, in a new take: the original is left as it was. The sung melody moves to an instrument, only the score's section tags are sent to the model in place of the words, the instrumental LoRA is used, and the style loses the tags that describe a voice ("soft male vocal"). It is the way to get "this tune, no vocals". Emptying the Vocal part of a song's score by hand is not enough: a song's render still sings the words it is given.
The Vocal chips set the singer's sex and character by writing into the style, for the whole song.
For a duet, or one singer for the verses and another for the chorus, put voice cues in the lyrics and tick Voices follow the cues: see The words under writing a song from a prompt. It works for a cover's words too.
A corpus is a folder of songs, prepared as a training set for a style LoRA: typically one
artist, one genre, or a few similar artists.
Open Corpora from the menu. It is on by default; TRAINING_ENABLED=0 for the app, or an
engine built with WITH_TRAINER=0, takes it out.
Training is the most demanding thing the app does: a 16 GB card is the practical minimum.
- New corpus. Give it a name and a trigger word, say whether the voice is male or female
(none for songs without vocals: see below), describe the sound shared by every song, and open the folder that holds the songs. Confirm you
have the right to train on them, then press Scan the folder. The folder is only read.
A corpus without vocals. Choose none for Voice when the songs are instrumental, such as
dance and electronic music. The app then does not separate a vocal or listen for words, so nothing
invents words over synths. Each song's lyrics are the section tags found in its score (or
[instrumental]when it has none), the caption leaves out the vocal clause, and there is no lyrics review: the Sections box can be edited, and an edit is kept. The sound of each song is described by the local model when it is installed, since the external one would only have a title to go on. A LoRA trained this way may write notes in the Vocal part of a plan, which makes a render sing or hum; a plan's "may sing" warning shows it. Changing the Voice of a corpus already analysed changes its lyrics to match. - Choose the songs. Untick any you want left out. A recording longer than 10 minutes is
left out, since it is most likely a whole album in one file. If a
.cuesheet sits beside it, Split into tracks cuts it into its songs, which take its place in the corpus. The tracks go in the app's own folder; yours is not changed. - Analyse. Each song's vocal is separated, its key, tempo and sections are found, and its lyrics are drafted, tagged by section. With an external LLM set in Settings, the sections are marked from the words as heard: choruses by their words coming back, new sections by the pauses. Only the lines and their timings are sent, and the words and their order are never changed. Without one, or if its answer does not hold every line, the tags come from the music analysis. Redraft, beside Save in a song's review, marks the sections again the same way without hearing the song again, for a song drafted before, and replaces what is in the box. A line under the buttons says what is running, and Stop ends it. Finished steps are kept, so Analyse carries on from where it stopped.
- Review. Open a song to check its lyrics and tick checked, and to describe its sound where it differs from the rest. The style caption shows what the trainer will read.
- Export training set. Writes the audio, lyrics and caption for each song. A line under the
buttons shows how far it has got. The trainer takes only so much of a song, which the app works
out from your graphics card's free memory (about 5½ minutes on a 16 GB card; the Logs window
says what it chose). A longer song is cut at the end of its last section within that, faded
out, and keeps only the words still sung. Its row says "first 3:12 trained". If training
runs out of memory, set the
TRAIN_MAX_MINUTESenvironment setting to a lower number, such as 3.5: see the README. - Train a LoRA. This takes a long time, and the GPU is not available to the app until it finishes. Progress shows under the buttons and on the main screen, where you can stop it, and the corpus's badge at the top of the page pulses while it trains. Training again keeps the LoRA from the last run under a dated name, in Previous runs, or deletes it, as you choose. Training needs at least about 12.5 GB of GPU memory (at a 3½ minute cut; a longer cut takes more), most of it while it prepares the songs, so a 16 GB card is the practical minimum; a smaller one, such as 8 GB, can't train, though it can still use LoRAs trained elsewhere (Install a LoRA). The Train window warns when less than that is free, so close anything else using the GPU first.
A run can end before its last step: Stop, a crash, or the GPU running out of memory. The line under the buttons then says when and why, and after which step. Along the way the trainer saves a checkpoint every 50 steps, and keeps a copy of the best one so far by its planner loss, so most of the work is usually still there. Finish with what it saved does what a finished run does, from those files: the best copy, or failing that the last checkpoint, becomes the LoRA under the corpus's name, with its trigger word, and the checkpoints fold under it in the Style LoRA list. It takes seconds and no GPU time. The last steps of a run change the LoRA least, so one finished this way is usually close to a full run; Checkpoints renders a song once on each step if you want to compare by ear. To train the whole run again instead, press Train a LoRA.
Run all, beside the three, does Analyse, Export and Train one after the other, for a corpus you would rather not wait on: a big one takes hours to analyse. It asks first what to do with an earlier LoRA, then carries on with the page closed. It analyses only what still needs it, and a song whose analysis fails is left out of the training set and named under the buttons. Lyrics nobody has checked are used as drafted. Stop ends it where it is, keeping what is finished; stopped while it exports, it finishes the training set first. It waits for anything already using the engine before it trains. An app restart ends it, and Run all again carries on.
A job that runs out of GPU memory says so, with what to try: close other programs using the GPU, lower the length cap, or, for training, leave the longest songs out. If the engine crashes and restarts, the job it was running is lost, and it says that too.
When training finishes, the LoRA appears in the Style LoRA list with its trigger word. Choosing it shows a chip for each song it learned from, under the Style box, and clicking one puts that song's style in the box. Past a dozen songs, a filter box and a row of the tags that recur across the corpus (its genres, moods and instruments) narrow them; click two tags to find songs with both. The list shows its first dozen until Show all. See Balancing Planner and Sound below for starting strengths. To share it, press Download LoRA on the corpus's page, or Download under the picker; see Sharing a LoRA below.
To add a LoRA someone shared, press Import LoRA beside New corpus and choose their zip. To use a LoRA trained elsewhere from the exported set, press Install a LoRA and choose the file. It is added to the Style LoRA list with this corpus's trigger word.
Deleting a corpus removes the app's copies of its songs, the separated vocals, the lyrics and the scores. The folder you pointed it at, and any LoRA made from it, are not touched.
What you train on, and what you do with the result, is your responsibility under the law where you live. The app asks you to declare it and does not check it: creating a corpus requires you to confirm that you have the right to train on those recordings.
Two facts are worth knowing:
- The YuE2 weights are CC BY-NC 4.0 — non-commercial use, whatever you train from them.
- Training privately on a corpus and publishing the LoRA or its output are different acts. The second is the one that usually needs permission: from whoever holds the rights in the recordings, and for a recognisable voice, from the singer.
LoRAs you did not train yourself carry their own licences: read them on the page each one comes from (see Where to find style LoRAs under Style LoRAs).
A LoRA is a small file that leans the model towards a sound. Put one in models/loras/ and press
Rescan under the picker, or use Install, and it appears in the Style LoRA list.
A style LoRA can hold two halves, and the picker shows which ones a file holds:
- Planner shapes what is played — the score plan (form, harmony, phrasing), and then the music the render writes from that score. It applies at both steps.
- Sound shapes the audio — timbre and production.
A strength the file cannot use is greyed out, and a file this engine cannot load at all is named as such rather than failing quietly inside a render.
The quickest way is at install: tick Style LoRAs on the installer's components page (or, with Docker, run
sh scripts/fetch-models.sh --loras). That fetches 39 style LoRAs by their authors, 3.8 GB in all, each with a
description giving its trigger word and suggested strengths: the sets below, two from atomtanstudio (dream pop
and old school hip-hop), and sixteen concept sliders from ntc-ai, which shape the score only (Sound does nothing,
and they need no trigger word). They are gathered, unchanged, in
pshields/yue2-style-loras, whose README credits each author
and links the page it came from. A LoRA already in your folder under the same name is left as it is.
There are more on Hugging Face. These are the sets gathered there, each on its author's own page:
- Militant reggae
- Chanson française
- Canzone italiana
- Bulgarian voices
- Qawwali, sufi and tabla
- J-pop, a single LoRA with a Sound half only
To add one, download its .safetensors file from the repository's Files tab, put it in models/loras/
and press Rescan, or press Install and choose the file. The repository's page gives each LoRA's
trigger word and suggested strengths: write them in a .txt beside the file (see Descriptions below) and
the picker shows them, and puts the trigger word at the front of the Style for you. A LoRA from elsewhere
may be saved in a layout this engine cannot load: the picker names it as such, and it does no harm.
Each LoRA carries its author's licence, which these sets give as CC BY-NC 4.0: personal and non-commercial use, the same as YuE2's own weights. Read the licence on a LoRA's page before using it, or sharing what it makes.
- Cohesive corpora: A LoRA trained on a single album or unified acoustic sound (e.g. 1960s folk rock) can run higher strengths, typically around Planner ~0.85 / Sound ~0.80.
- Diverse corpora: If the training corpus spans multiple genres, production styles, or eras (e.g. acoustic folk, rock, and synth-pop), high Sound weights can cause acoustic clashing. Up to about Planner 0.70 / Sound 0.70 keeps the audio clean while retaining the artist's melodic phrasing and vocal character.
- Covers: your recording sets the melody, so there is less for the LoRA to shape and the Sound half is pushed harder. Keep Sound near 0.50.
- Plan variety: with a LoRA trained from a corpus, Calm or Normal gives the most recognisable result. Calm can loop with a style LoRA and a short style, and the plan then runs for minutes with long repeated intros, or with nothing sung. Describe the style in more detail (genre, instruments and feel), or use Normal. A plan like that, or one far longer than a song, is written once more with a new seed, and the take fails with advice if the second comes out the same.
- Finding a good take: the LoRA's character depends on the roll as much as on the strengths. Between about 0.3 and 0.6, Sound changes the voice little, so there is no need to push it. Planner changes the performance itself, not only the voice, so a value that suits one song can be the weakest for another. Treat it like the seed: try a few (say 0.6, 0.8 and 1.0) and a few seeds with Sing again, and keep the take you like.
- Save strengths: when you find the right pair for a LoRA, press Save strengths under the picker. Choosing that LoRA then starts at them, and they travel with it when you share it.
Most style LoRAs are trained on captions that begin with a trigger word, and do very little without it. The app handles this: choosing a LoRA puts its trigger at the front of the Style, changing to another swaps it, choosing None removes it, and a render puts it back if it was deleted. You will see the word appear in the Style box — it is yours to edit or move.
The list is grouped by set. Click a heading to open or fold it; the picker remembers which you left open, and the group holding your current choice always opens with it.
Hover any entry to read what it is, in its author's words, with their suggested strengths. Those descriptions come from a text file beside the LoRA:
CHNSN Rive Gauche
Trigger: chnsn
Step 200, the decoder-loss minimum. The more supple of the two: acoustic
narrative, yé-yé, female leads, waltz meters.
The first line names it, a Trigger: line becomes the trigger word, and the rest is the
description. A LoRA of your own gets one by writing a .txt beside it. families.txt in the same
folder gives the groups their headings, one prefix = label per line.
A LoRA trained from one of your corpora is an ordinary style LoRA: it appears in this list with the rest, its trigger word beside it, and the same two strengths apply.
A training run saves a checkpoint every 50 steps, and keeps them under its LoRA in the picker: ▸ 9 steps beside the LoRA's name opens them (Settings can delete them instead, to save the space). A checkpoint whose LoRA has been deleted stays under Training checkpoints. To clear some out, press Edit on the corpus: Training checkpoints there lists each run's, this run's and any previous run's, with their sizes. Tick the ones to go and press Delete. The finished LoRA is not listed; it has Delete LoRA of its own. Each is the LoRA as it stood at that point in training. They sound about as good as each other, but each has its own weighting of what it learned, so each has its own taste while keeping the style and signature sound the LoRA was trained on.
To hear them side by side, choose the LoRA, set up the editor as you would for one take, and press Checkpoints beside Delete LoRA. It is greyed out for a LoRA without checkpoints. Tick the steps you want, set a length cap for these takes if you like, and press Render. The editor's own action runs once on each checkpoint, all with one seed, and each take is named after its step: Night drive · step 250, Night drive · full. They are kept inside the take open in the editor, or, for a new song, inside the last of them (see Takes inside a take, under The library). The editor closes, so you are back at the library where the takes are being made. A song or an instrumental writes its plan and goes straight on to render.
In a song or an instrumental the LoRA writes the tune as well as shaping the sound, so the same seed does not give the same song: each checkpoint writes its own melody and structure. The seed picks from what the model thinks likely, and each checkpoint thinks slightly differently from the first note, so the plans part ways within a few bars. What stays is the sound. To compare only the sound, load one plan with Score and press Render this score with each checkpoint in turn.
New files appear once the engine has looked at models/loras/ again, which it does when the
options are reloaded.
Choose a LoRA and press Download under the picker: you get one zip with the LoRA and its note. For a LoRA trained from a corpus, the note carries its learned styles, so the chips come with it.
To add one someone sent you, press Install and choose their zip, or a bare .safetensors
file. It goes into models/loras under the heading Installed, with its name, trigger word and
chips. Delete LoRA removes the chosen LoRA, its note and its group line; Rescan reads the folder
again after you add a file by hand.
Press Stems on any take. Choose a model, tick the parts you want — vocals, drums, bass, other, and guitar and piano on some models — and run.
Separation runs on the GPU when the engine has Demucs, as a job in the engine's queue: it waits for a
render that is running, then goes quickly, and a render queued behind it waits in turn. An engine built before
this, a LoRA being trained (which holds the GPU), Use the GPU for stems and lyrics set to Off in Settings, or STEMS_ON_GPU=0 leaves it on the CPU, which is slower
and never holds up a render. Fine tuned runs four models in turn for a better split and takes about four times
as long.
For a backing track, choose Vocals and instruments. It splits the song in two: the vocal, and instruments, everything else as one stem. Untick vocals to keep only the instruments. It is the same separation as the four stems, added back together, so it takes as long and sounds as clean. The fine tuned one leaves the least of the voice behind.
Stems land in data/stems/<title>-<id>/, play from the chips on the card, and download singly or as
a zip. They stay until you delete them.
A queued or running take shows Cancel on its card, and stop on the job card stops whatever the engine is working on. Deleting a take stops its job and removes its stems. Delete beside a recording removes the file and its stems; covers made from it keep their audio and score, but cannot be rendered again.
+ Song, + Cover and + Instrumental on the take panel, or New song beside the editor's heading, start again from the take on show. They clear the title, the lyrics and the score, and start from the defaults: an empty style (with a chosen LoRA's trigger word), the usual length cap, Harmony Varied, plan variety normal, interpretation Wide, production polish on, normalising off, and a new seed that is not held. A chosen LoRA and a chosen recording stay chosen. The take being shown lets go of the editor, so Render cannot act on it by mistake.
Words you had typed but not used are not thrown away: a bar offers to restore them.
Takes live in spaces: a space per song, per album, or for sketches. The menu above the takes chooses which is on show, and anything you make lands there. Deleting a space never deletes takes — they move to Default. Recordings are shared by every space, so one recording can be covered in several.
The folder button on a card moves that take to another space, with any takes inside it.
Variations, Try more and Checkpoints each make several takes of one song. They are kept inside the take they were made from, so the space lists your songs and not every render of each. Checkpoints is pressed in the editor: with a take open there, its takes go inside that take; for a new song, which has no take yet, the take on the last checkpoint ticked (the finished LoRA, when it is one of them) holds the others.
- The Inside button appears among the small buttons at the top of a card once its take holds takes, finished or still being made, and it is lit orange. Hover over it to read how many are done and how many are to come. The card itself is plainly lighter than the others (in the light theme, a warm tint), so a take with takes inside it stands out in a list. (The mastering rack's button beside it is lit green in the same way, when the take is mastered.)
- Click it and a panel opens under the card with the takes inside, each a card of its own. A take shows there when it is finished; while it is being made only the take it belongs to shows. Click the button again, or Close, to shut it.
- A take inside another is a take like any other. Play it, star it, save it, split it into stems, rename it, and double-click it to open its song and score.
- One level. A take inside another holds none of its own, so its card has no Variations or Try more, and Checkpoints in the editor is greyed out while it is the take open there. Take out (the arrow leaving a box) makes it a take of its own in the space's list, and then it can.
- The list's tools follow the panel. While one is open, Select all, Starred, the kind buttons, Compare, Delete and the player's previous and next work on the takes in the panel. The take whose panel is open stays on the page whatever the filters say.
- Something playing inside is shown outside. If you shut the panel while a take in it is playing, the Play button on the take that holds it glows gold, so you can see where the sound is coming from. Green still means that take itself is playing.
- Search finds them like any other take, wherever they are. While a search is on, what it finds is shown as plain cards, and a take that is inside another says which.
- Deleting a take deletes the takes inside it. The warning says how many. Take out any you want to keep first.
Takes made before this existed stay where they are, side by side in the list.
Each card names the settings that shaped it — Harmony, Interpretation, Plan variety, the LoRA and its strengths, the seed and its age — so a card reads as the recipe that made it.
Save asks which format to download the take in, FLAC, WAV or MP3, starting with the output audio format set in Settings.
Search takes finds the takes with every word you type somewhere in the title, style, lyrics or LoRA name, in any case. It looks through the space on show; All spaces, beside it once you type, looks through every space, and a card from another space says which one. × or Escape clears it. The box starts empty each time the page opens, and offers your last few searches that found something: click one, or pick it with the arrow keys and Enter.
Starred shows only starred takes. Beside it, two buttons name what a click does: Comfy or Compact switches between wider cards with more of the prompt and three narrow cards across, and Fit or Wide between a centred column and the whole width of the window. A title always stays on one line, ending in … when it is long. Hover over a card's title, settings or style to read all of it.
Each new take gets a picture of the song, shown faintly behind the take's card, across the top of the take panel when you choose it, and behind the editor when the take is open there. Drawn on this PC it waits until no song is being made, so it never holds up a plan or a render, and appears a little after the song is finished; drawn by an external LLM it is made while the take renders. It is made in two steps: the take's title, style and opening words are turned into a description of a picture, then an image model draws it on white, and the white is taken out, leaving the picture with a transparent background. A take made inside another (Try more, Variations, Checkpoints) shows the picture of the take it is inside rather than having one of its own.
An image model will not draw a real band or person, so the description leaves names out; and the picture's own lettering is kept soft, so it does not fight the card's.
Pictures need something that can draw. There are two, and you can have either, both or neither.
| FLUX.2 klein, on this PC | Your external LLM's image model | |
|---|---|---|
| What it is | a small image model that runs on the engine, beside YuE2 | the provider's own image model, such as Gemini's |
| What it costs | nothing; about 8 GB of disk | each picture is one request to the provider, which may be charged |
| What leaves the PC | nothing: the description is made from the take itself, with no LLM | the take's title, style and opening words, and the description |
| When it runs | on the GPU once no song is being made: after your song, never in its way | beside the renders, while they run |
| What you set up | an install option | the external LLM in Settings, with a provider that lists an image model |
To draw on this PC, install FLUX.2 klein.
- Windows: tick FLUX.2 [klein] on the installer's components page. It is not ticked unless you choose it. An installed app gets the same page when it updates, or run the installer again from the release page: the tick is remembered after that, and an update never downloads it twice. It is three files, about 8 GB in all.
- Docker: run
sh scripts/fetch-models.sh --pictures, which puts the three files inmodels/(diffusion_models,text_encodersandvae) and checks each one, then restart the app withdocker compose restart appso it finds them.
To draw with an external LLM, open Settings and set LLM for writing lyrics & style tags to the
external LLM, with its API URL, API Key and Model, as for lyrics. The provider must have an image
model: Gemini does (gemini-3.1-flash-image is used unless you name another), and an OpenAI-compatible
provider is asked for gpt-image-1. The app asks the provider for its list of models, in the background,
when the app starts, when these settings change and every few hours, and draws there only once the list
has the image model.
With neither, Take pictures in Settings says No pictures: the default background, with a line on how to have them, and the choice to make them cannot be picked. Set up either and they come on: pictures are then made for the takes you make from then on.
| Setting | Does |
|---|---|
| Take pictures | Make one for each new take (the default where something can draw), or No pictures: the default background, where none is made unless you ask for one in the editor. It is No pictures while nothing here can draw. |
| Take pictures: drawn by | This PC if FLUX.2 klein is installed, else the external LLM (the default), This PC (FLUX.2 klein) or The external LLM's image model. |
| Take pictures: described by | From the take itself (the default): made from the take's title, style and opening words with no LLM, whoever draws; the quickest, and nothing is sent out for it. Or Local Gemma, whose descriptions are more inventive but take far longer and need the Lyric drafts option, or the External LLM, a request each. |
| Take pictures: image model | which of the provider's models draws. Only for pictures drawn by the external LLM (choose it under Drawn by; greyed out otherwise), and empty for the provider's usual one, which is nearly always what you want. |
So with FLUX.2 klein installed and the settings as they come, pictures are made entirely on this PC, even when the external LLM writes your lyrics, and they need no LLM at all: a description made from the take takes no time, so a picture is ready a few seconds after its song. Choose The external LLM's image model to have the provider draw them instead, or set described by to Local Gemma or External LLM for their descriptions over the take's own, at the cost of time (and, for the provider, a request each).
The Picture icon in the editor window's top row, beside the take's other buttons, shows the picture with two choices: Make a new picture, which makes it again in the background, and Use the default background, which takes it away for good: no picture is made for that take again unless you ask for one. Click the picture there to see the whole of it, as large as the window allows; a click anywhere or Escape puts it away and leaves the editor as it was.
Tick two to six takes and press Compare (it appears beside Select all once two are ticked, and waits for takes still rendering). The takes stack in one window on a shared position: press 1 to 6 or click one to switch, and it carries on from the same moment in the song, playing or paused. Space plays and pauses, the arrows move 5 seconds, and the star marks the one you prefer. Match loudness turns the louder takes down to the quietest, since a louder take tends to sound better; Hide names shuffles the takes and calls them Take A, B, C for a blind listen.
Click or drag the waveform to seek. Previous and next step through the cards in the order shown.
| Key | Does |
|---|---|
| Space | play or pause |
| Left, Right | back or forward five seconds |
| Up, Down | volume |
The system media keys work too. The take's own keys are under Hot keys.
On the main page, single keys act on the chosen take: the one in the left panel. Click its card to choose it. Each key does what the button for it does, so a key that would delete asks first, and a take that cannot do what a key asks (Variations on a take with no score, say) leaves a line in the status bar saying why.
| Key | Does |
|---|---|
| Space | play or pause |
| Left, Right | back or forward five seconds |
| Up, Down | volume |
| M | open the mastering rack for the take; again to close it |
| E | open the take in the editor |
| I | the take's details: everything it was made with, in words, with Copy |
| N | normalise the take, or undo it if it is normalised |
| V | Variations |
| T | Try more |
| S | star the take, or unstar it |
| D | delete the take, after asking, as its bin does |
| L | the logs; again to close them |
| ? | show or hide the hot keys; again to close |
| Escape | close the window in front: the logs first, then the mastering rack over the editor, then the editor or any other window |
The letter keys do nothing while you are typing in a box, while a window is open in front, or with Ctrl, Alt or the Windows or Command key held, so the browser's own shortcuts (Ctrl+L, Ctrl+S and the rest) still work. A button's tip names its key, as in Variations (V). Press ? at any time to display the list of hot keys.
In the editor window, Space and the arrows work the window's player, for the take open there, whenever its player shows: on every step, and with the mastering rack open over it. They never act while you are typing in a box, and the letter keys are the main page's only. The piano roll has keys of its own: see The Piano Roll.
Press Yeufonic in the top left. Settings live on the server, so they follow you to any browser and survive a rebuild: the output audio format for stems and a take's Save, the separation model, where stems are written, whether training checkpoints are kept, and how instrumentals are checked for singing.
Settings → Storage shows where Yeufonic's disk space is going and lets you give some back. Training a corpus leaves a lot of files that are not needed again: copies made for the engine while songs were analysed, the training set (each song written again as lossless FLAC, larger than the original), and the checkpoints saved along the way. The window lists each of these with its size, what it is and what removing it costs, and removes only what you tick, after a last confirmation. Tick the ones that cost nothing ticks the copies that are simply made again when needed.
Songs analysed from now on keep their separated vocal as FLAC, the same audio in about half the room. The window can convert the WAV vocals you already have: each is converted, checked to decode to exactly the same audio, and only then is the WAV removed. It runs in the background and can be stopped.
Training checkpoints deserve a second look before they go: each is an entry in the Style LoRA list that you can pick for new takes, so removing a corpus's checkpoints takes those entries out of the list for good. Its finished LoRA stays and keeps working. A corpus's training set is different: removing it leaves the LoRA and its checkpoints in the list, and only training that corpus again needs it.
Your takes, songs, finished LoRAs and models are never listed for removal. Anything belonging to a corpus that is training, being exported or being analysed is shown greyed and cannot be touched until it has finished.
Below that, Tidy up automatically chooses what Yeufonic removes for you: the working copies of each song once its analysis ends (on by default), the training set once a LoRA has trained (off by default; the corpus then needs Export again before it can train again), and the training checkpoints (kept by default).
Docker's own images and build cache are outside Yeufonic and not covered here; docker system df shows them.
YuE2 comes in two sizes, and an install has one: full quality, which we recommend, or low memory, a smaller copy for graphics cards with little video memory. About, in the same menu, says which you have. It is chosen when Yeufonic is installed, not here:
- On Windows, run the installer again. Its components page offers both sizes, shows the one you have, and switching downloads the other and removes the first.
- With Docker,
sh scripts/fetch-models.shfetches full quality andsh scripts/fetch-models.sh --int8the low-memory one. With both files inmodels/checkpoints, the app uses full quality unlessYUE2_CHECKPOINT=int8is set in.env.
Try low memory if renders are very slow or the engine runs out of GPU memory. Training a LoRA uses the full-quality model and says so if only the small one is installed. The README on GitHub has a section, Choosing a model size, with what to expect from each.
Check for a new version is the one thing here that uses the internet on its own account: every four hours the app asks yeufonic.com whether a later release is out, and says so with a pill beside the version in the top right. That pill hands you the installer, or, for a Docker copy, the two commands that update it. It is one request and nothing else leaves the computer; Do not check turns it off. The menu's Check for updates asks then and there whether this is on or off, and answers in the menu: which version you are on, and whether it is the latest.
When there is a newer one, the menu offers Get x.y.z, which downloads the installer on a Windows install or puts the two commands on the clipboard for a Docker copy.
Theme sets how the app looks: Dark, Light, Match the computer, which follows your system's light or dark and changes with it, Studio, a warm dark with amber, or High contrast, black and white with strong outlines. It changes at once, and each browser remembers it so the page opens in it. The log window stays dark in every theme, like a terminal. The mastering rack follows the theme too: in the light theme its panels are brushed silver, while its knobs, meter windows and readouts keep the dark look of the hardware.
Backdrop colour, under Theme, tints the page behind the takes with a colour you choose: tick Tint, then move Hue round the colour wheel and Strength from a faint wash to a rich colour. The page changes as you drag, and the choice is kept when you let go. The take cards, the panels and the windows keep their own colours, and the high contrast theme stays black.
Yeufonic can act as an MCP server. That lets an AI agent on the same computer, such as Claude Code or Cursor, use it for you: ask for a song, a cover or an instrumental, wait for it, play it, and tidy the library, all in your own words ("make a one-minute EDM instrumental in the EDM space and play it"). The agent does what the page does, through the same routes, so a take it starts is an ordinary take in your library.
-
Turn it on. Open Settings and set MCP server to On. It is off by default, because a request can start work on the GPU.
-
Connect your agent to the app's address with
/mcpon the end: the address you open Yeufonic at, for examplehttp://localhost:8090/mcp. In Claude Code:claude mcp add --transport http yeufonic http://localhost:8090/mcpOther clients take the same address as an HTTP server. Those that read a settings file usually want something like
{"mcpServers": {"yeufonic": {"url": "http://localhost:8090/mcp"}}}. No login or key is needed. -
Ask. Start your agent from any folder and describe what you want.
Which address? It is the address you open Yeufonic at, with /mcp on the end.
- Docker:
http://localhost:8090/mcp. - The Windows install: the app's port is in
settings.ini(app_port, 8090 unless you changed it, for instance to run beside a Docker copy). Withapp_port = 8091it ishttp://localhost:8091/mcp. - Agent in WSL, app on Windows: this does not connect. The Windows app listens only on Windows' own loopback address, which WSL cannot
reach (Docker running inside WSL is fine, as it is in WSL too). Run the agent from Windows (PowerShell or the Windows version of
your agent), or turn on WSL's mirrored networking (
networkingMode=mirroredunder[wsl2]in.wslconfig, thenwsl --shutdown), which makeslocalhostthe same on both sides.
If your agent says the server needs authentication, it is almost always because the setting is still Off: turn it on, then
reconnect the agent (in Claude Code, /mcp). Reconnect after restarting Yeufonic too, because a restart ends the connection.
| Ask for | What happens |
|---|---|
| "List my latest takes", "show the takes in the EDM space" | Lists takes, newest first, with a short id, when each was made, its length and style |
| "Make a one-minute instrumental in a synthwave style" | Plans and renders an instrumental. A length, a space and a style LoRA can be named |
| "Make a song in a folk style with these words" | The same for a song, from the lyrics you give |
| "Make a cover of the recording in a jazz style" | Makes a cover from a recording already in your library. It needs a score (transcribe it in the app first). Without words of your own it uses the words heard in the recording |
| "Play it" | Opens the take in your browser, where it plays |
| "Star that one", "call it Sunrise", "move it to the Folk space" | Stars, renames or moves a take |
| "Render it again with a new seed", "stop that" | Renders a take again, or cancels one in progress |
| "Give me a few more like that one" | Renders the take again with new seeds, as several new takes to listen to together |
| "Transcribe the recording" | Transcribes a recording into a score, so a cover can be made from it. It leaves a recording that already has a score alone unless asked |
| "Split that into vocals and instruments" | Separates a finished take into its parts (four, six, or just vocals and instruments) and gives you a link to each file and one to the whole set as a zip |
| "Is Yeufonic busy?" | Says whether the engine is ready, what is running and with what progress, what is waiting, and whether a LoRA is training |
| "Delete that take" | Shows what would go and asks you first; see below |
Making a take takes minutes. The agent starts it and waits, reporting progress, so you can ask for the next thing meanwhile.
You do not need these names: you ask in your own words and the agent picks. They are here so you can see what exists, and so you can tell your agent which one you mean.
| Tool | What it does |
|---|---|
list_takes |
Lists takes, newest first, optionally only those matching words, in a named space, or starred. Shows a short id |
get_take |
One take in full: how it was made, its status, any error and, once it has audio, a link that plays it. Can include the score |
wait_for_take |
Waits (up to about a minute at a time) for a take that is being made, and says whether it has finished |
make_instrumental |
Plans and renders an instrumental from a style, with a length, a space and a style LoRA if you name them, and sections if you give them |
make_song |
The same for a song, from lyrics |
make_cover |
A cover from a recording in the library, using its score and, unless you give others, its words |
try_more |
Several more takes from an existing one, with new seeds |
render_take |
Renders a planned take, or renders a finished one again (with a new seed if asked) |
cancel_take |
Stops a take that is queued or running |
make_stems, get_stems |
Splits a take into parts, then lists the parts with a link to each and to the whole set as a zip |
list_recordings, transcribe_recording |
Shows the recordings in the library and whether each has a score; transcribes one that has none |
list_style_loras |
The style LoRAs installed, with their trigger words and the strengths saved with them |
list_spaces, move_take |
The spaces and what each holds; moves a take to another |
star_take, rename_take |
Stars or unstars a take; changes its title |
delete_take |
Deletes a take for good, after showing it to you and being told to go ahead |
status |
Whether the engine is ready, what is running and waiting, and whether a LoRA is training |
- A batch of ideas, then pick. "Make five one-minute EDM instrumentals in different styles, in the EDM space." It starts them one after another and tells you as each finishes. You listen, then say "star the second and fourth and delete the rest."
- Iterating on one idea. "Make an instrumental in a dreamy synth style and play it", then "now the same with more drums", and "give me four more like the second one." The agent keeps track of which take is which, so you can talk about them by position or by what you said about them.
- Covers without clicking through. "Transcribe the demo, then make three covers of it: folk, synthwave and jazz." It transcribes the recording first if it has no score, then makes the covers in the space you name.
- Working a LoRA. "List my style LoRAs, then make an instrumental with the jazz one." The trigger word and the saved strengths are handled for you, which is easy to get wrong by hand.
- Words from the agent. The agent can write the lyrics itself: "Write lyrics about a night train and make a song of them in a folk style." Yeufonic only sees the finished words.
- Stems for a whole set. "Split my five newest takes into vocals and instruments and give me the links." They queue behind one another, and behind any render the GPU is running.
- Library housekeeping. "Find the takes with the same title, show me which is newest, and delete the older ones", or "move everything I made today into a space called Drafts." It works from ids and asks before deleting.
- Checking before a long job. "Is Yeufonic busy?" before you queue a long render or a batch, so you do not queue behind a training run.
- Part of a bigger task. Because it is a tool the agent can call, it can sit inside something larger: make a track for a video and write the note that goes with it, or build a playlist-sized batch and name the takes to match a list.
The agent cannot hear a take, so it cannot judge how it sounds: you do the listening, and it does the legwork. It is also not a shortcut round the engine: the GPU still does one job at a time, so a large batch is a long queue.
- Takes can share a title, so the agent works from a take's id and shows a short one (the
short_idcolumn) when it lists takes. You can name a take by the start of its id, four characters or more, as long as only one take begins that way. If two do, it tells you and asks for more. - Deleting asks twice. The agent first shows you what it would delete, including any other take with the same title, and deletes only after you say yes. A starred take needs you to say so explicitly. A deleted take, with its audio, cannot be got back.
- Spaces and style LoRAs are named, not numbered: "in the EDM space", "using the jazz LoRA". A name that matches nothing gets the list of what there is, and nothing is created. A LoRA's trigger word is added to the style for you, since a LoRA does very little without it.
- An instrumental is always planned with a structure that has times, worked out from the length you ask for (about two and a half minutes when you do not say). Without times the planner writes plans too long to use. Name your own sections and they are used as given.
- Only on this computer. The same host-name and cross-site checks apply as for the page, and there is no login, so turn the setting off when you are not using it, and do not expose the port to a network you do not trust.
The app writes a consolidated, real-time log of every major action — score planning, rendering, audio transcription, stem separation, and LoRA training — tagged with INFO, WARN, and ERROR prefixes.
- In-Browser Console: Click the Logs button in the topbar (or select Logs from the brand menu) to open a floating, draggable, and resizable console. It stays open without blocking the page, allowing you to queue takes and monitor generation in real time.
- Standalone Window: Click ↗ Pop out in the console header (or navigate directly to
/logs) to open a dedicated log viewer window. - Terminal Tail: Logs rotate automatically to
data/logs/yeufonic.logon the host, where you can follow them live:tail -f data/logs/yeufonic.log
The engine is ComfyUI, and its own interface is reachable: http://localhost:8189 with Docker, or
http://localhost:8188 with the Windows install (the engine_port in settings.ini). It answers
on this machine only, with no login.
- It is the same engine and GPU as the app. A job run there shows in the app's queue as coming from outside, and the app's own jobs wait for it.
- Its History shows every graph the app sent: the plan, the render and the transcriptions, with every setting as the engine received it. It's a useful place to see exactly what a take was made from.
The header says the engine is missing. The app runs without it and will say so. Renders wait. Check the engine container is up.
A song will not end. Usually too much planner strength. Drop the LoRA's Planner to 0.5. The length cap will stop it regardless.
A LoRA seems to do nothing. Check the trigger word is in the Style, and that the strength you raised is one the file actually holds — the picker greys out the other.
A take says Weak render. It came out far quieter than usual all the way through. Some takes like that are only quiet; others sound thin or distorted. Click the warning to normalise it, which brings it up to the usual loudness and marks it Normalised. If it still sounds wrong, try another seed, and keep that LoRA's Sound at 0.5 or below. It happens most in covers through a style LoRA trained here.
Normalise volume. Tick it in the editor and each take made while it is ticked is brought to the usual loudness when it finishes, and marked Normalised. A take made without it can be normalised later with the speaker button at the top of its card. The file as rendered is kept beside it, and clicking Normalised goes back to it. How loud it is made is set in Settings, under Normalise to. Unticked, a take keeps the level it was rendered at.
An instrumental sang. See When an instrumental sings above. It is a model failure; a new seed usually fixes it.
A take will not render again. Its recording may have been deleted. Covers keep their audio and score, but cannot be re-rendered from a recording that is gone.
Nothing plays, but the waveform moves. Check the volume slider and that a stem is not selected — the player follows whatever was last clicked.