Add songs from My Songs (๏ผ Playlist on a song). โถ plays a playlist in the player at the bottom; tap a song to start from there.
Paste a Spotify track, playlist, or album URL โ or a YouTube video or playlist link (youtu.be share links and full youtube.com URLs both work; saved as MP3 with the video thumbnail as cover art) โ to download it and use as a cover source. A playlist or album downloads every track in it. You can also upload audio files of your own. Saved tracks (and their transcribed melody, once transcribed) live here so you can reuse them without re-downloading or re-transcribing โ pick "Cover" on one to jump to Create with it already loaded.
Record a clip or save a photo, then turn it into a style + lyrics suggestion โ optionally with a hint about what you want, e.g. "Make a comedic song about this character taking a hot bath in nacho cheese." Whatever you create from lands back on Create, ready to generate.
Recording stages the clip on this device first, so a dropped connection during upload never loses the take; it just retries.
Everything you've saved and how much space it takes. Delete what you don't need to free up room.
You've reached your storage limit. New songs, uploads and downloads are paused until you free up space.
Everyone with an account on this instance. Tap someone to view their profile.
Loadingโฆ
Scores saved on their own, independent of whatever song or cover they may have started as โ pick one to load it back into Create, or preview it right here before you do.
Clone a voice from a short, clean clip of someone talking โ or design one from a description โ then have it say anything: intros, outros, interludes, ad-libs, skits. Add the result to a song or a Beat Maker project. This is speech, not singing (for singing in your own voice see My Voice). Powered by Breeze TTS 2 โ research and non-commercial use only. Only clone voices you have the right to use.
To chat with a voice or have two talk to each other, give it a character on the ๐ญ Actors screen.
An actor is a character with a voice: pick one of your spoken voices and write their lorebook in markdown โ who they are, how they talk, what they know. The lorebook is their system prompt when you talk to them and when the AI writes conversations for them.
Chat in character โ type, or tap ๐๏ธ and talk. Replies come back as text and in the actor's voice. "Keep as clip" saves a reply for songs.
Two actors talk to each other. Write the script yourself, or give them a topic and let the AI write it from their lorebooks. Each line is saved separately; play or export the whole thing as one track, or save it as a clip to add to a song.
Saved drafts of LoRA training data. Build one from My Cover Library (Select tracks โ ๐ Training set), then set a master style, override it per track, and check each track's lyrics before exporting a ZIP for AI-Toolkit or training here.
Automatic listening tests: the same lyrics across LoRA checkpoints and strengths, each scored for how much the singer sounds like the training set's voice. Start one from a training set (๐งช Evaluate LoRAs).
| # | LoRA | Planner | Decoder | Seed | Singer | vs baseline | Lyrics | Status |
|---|
Singer scores cluster near the top of the scale โ compare rows with each other and with the baseline rather than reading the number on its own. The top few are the ones worth a listen.
Tracks without lyrics are left out โ transcribe them first.
Runs in the trainer container and adds the finished LoRA to Style LoRAs. While it uses this server's GPU, new songs wait. Progress is under Admin โ LoRA training.
A beat project is a stack of tracks โ drum patterns, uploaded or recorded audio, even a short AI-generated clip โ each placed on its own timeline. Open a project to start stacking tracks, or start a new one below.
Loadingโฆ
Tap steps to build a beat โ drag across a row to paint several at once. Tap a bass or pluck step again to cycle its note.
Pick one of your spoken clips (make them on the ๐ฃ๏ธ Voices screen).
Pick one of your own generated songs or saved covers to use as a track.
Pick a saved score, or write one from scratch โ either way, preview and play it here before it's rendered to audio and added as a track.
Hum or sing the melody โ SheetSage2 transcribes it into an editable ABC score, the same way "Start from a recording" does on Create. No lyrics needed, just the tune.
Recordingโฆ tap Stop when done.
One free-text description, not a tag list โ YuE2 reads genre, instruments, vocal character, language, and tempo all together. Example: "English, warm piano pop, expressive female voice, 88 BPM." The Style preset dropdown above the field fills in a starting point you can edit. The model isn't limited to those examples โ it can also do electronic styles like techno, house, drum & bass, and happy hardcore; just describe the tempo (BPM), energy, and instrumentation you want (e.g. "aggressive happy hardcore, distorted four-on-the-floor kicks, euphoric synth stabs, pitched-up vocal chops, 170 BPM").
Use section tags โ [Verse], [Chorus], [Bridge],
[Intro], [Outro] โ each on its own line, followed by one sung line per
line. Only include words meant to be sung; stage directions or notes-to-self will get sung too.
See the project's own
full formatting guide โ for more detail.
Controls how much of the song's editable score YuE2 writes out before generating audio.
A number that determines the randomness used during generation. The same style, lyrics, and seed will always produce the same song. Change the seed to get a different take on the same lyrics/style; keep it the same while you tweak other settings to compare results fairly.
Classifier-free guidance scale โ how strictly generation follows your style/tags description versus how much freedom the model takes. Higher values stick closer to what you described (but can sound less natural if pushed too far); lower values give the model more creative freedom. YuE2's own default is 1.0 (a touch higher, 1.01, when plan mode is "off") โ "Use default" leaves it at that; the slider lets you try values from 0.5 to 3.0. Pushing it much higher isn't a guaranteed quality improvement, per YuE2's own docs, so the slider stays in the range worth actually experimenting in rather than the full 0โ20 the API technically accepts.
Upload, record from your microphone, or capture audio playing in another app/tab. SheetSage2 transcribes its melody into an editable ABC score, and Whisper transcribes its lyrics, so neither has to be typed by hand. The original recording is always saved alongside the song it produces, so you can go back and reuse it later. Use plan mode "melody" so the accompaniment can adapt to a new style around the transcribed melody. Recording from the browser needs this page served over HTTPS (or localhost). Models download on first use, which can take a while.
If plan mode wrote a score, it's rendered here as sheet music. Drag a note up or down to change its pitch, or click a note to jump to its position in the raw text below โ edits to either stay in sync. Paste in your own ABC score to generate from it directly.
Mixes a synthesized binaural beat underneath the generated song โ two pure sine tones a few Hz apart, one per ear, which the brain perceives as a "beat" at the difference frequency. This is plain signal synthesis, not something YuE2 itself generates (it has no way to hit an exact per-channel frequency). The "174Hz Healing" style preset fills in a commonly-used base frequency and switches this on; the base/beat preset dropdowns fill in the number fields below them, which you can then fine-tune directly. Needs headphones โ the stereo separation it relies on is lost on speakers.
Two different traditions are mixed into the presets, worth telling apart: the named "solfeggio" frequencies (174โ963 Hz) are a folk/wellness convention with no clinical evidence behind the specific numbers โ included because they're the frequencies people specifically ask for, not as a medical claim. 40 Hz gamma is different: there's real, if early and still unsettled, neuroscience research on rhythmic 40 Hz sensory stimulation and its effects on brain activity โ worth knowing that's a genuinely distinct (and more scientifically discussed) idea from solfeggio, not more proven for pain relief specifically.
Record or upload a few minutes of yourself talking or singing to build a reusable voice clone. Once it's ready, apply it to any generated song from that song's card โ the original song is never changed, so you always keep it alongside the version sung as you. Training runs on the same GPU as song generation, so it happens one at a time and can take several minutes.
Lets people sign in with Google, and request an account with it (requests still need your approval, marked "Google"). Create an OAuth client ID (type "Web application") in the Google Cloud console, add the redirect URI below as an authorized redirect URI, then paste the client ID and secret here. The Google button only shows once both are saved.
If songs fail with a CUDA out-of-memory error because another program was using the GPU, free this app's GPU memory. The song model unloads and reloads on the next generation, sized to whatever memory is free then. ComfyUI on the same GPU keeps its models loaded until told otherwise โ Free ComfyUI memory asks it to unload them (it reloads on its next workflow; a LoRA training run does this automatically).
Start a run from ๐ Training Sets (build one from My Cover Library: Select โ ๐ Training set). Finished LoRAs are added to Style LoRAs below automatically.
Which vision-language model captions photos for the Inspiration feature. Switching takes effect on the next photo captioning call, not immediately โ only one is kept loaded at a time.
Register a Hugging Face repo of style LoRAs trained with AI-Toolkit (or similar) so users can select one on the Create form. A repo can bundle several style files โ find them, then add whichever ones you want available.
Drag inside the box to move it, drag the corner handle to resize โ it always stays square.
5โ20 seconds of one person talking clearly works best. The transcript must match the clip word for word โ leave it blank and it's drafted automatically for you to check.
Describe who's talking โ age, character, tone, accent, pace.
Makes a new version of the song with the clip added; the original isn't changed.
Generates the lyrics below once per combination, then scores each result against this training set's voice. Everything runs on the normal queue; you can leave it and come back.
Training sets are saved drafts for a LoRA: pick tracks here, then set a master style, per-track styles, and check lyrics on the Training Sets screen before exporting or training.
Edited โ Save to keep these changes on the song itself, or Close to discard them.