Style Trainer setup
Requires full PyTorch YuE2 + VAE weights and a BF16-capable NVIDIA GPU. GGUF weights cannot train or use these LoRAs. Select local model paths in Advanced settings, then check setup. No community encoder or lyric files are needed.
Check existing files first. Downloads and GPU training only start after separate explicit actions. After a download finishes, click Check setup, then Use downloaded model paths.
Model terms: YuE2-3B license · VAE license. If FFmpeg is missing, use 48 kHz mono/stereo WAV or FLAC, or install FFmpeg on PATH for conversion.
Build your sound.
Choose your songs, share one style, and train an acoustic LoRA.
Trains acoustic production style from your recordings. It may influence vocal texture, but does not specifically train singer identity or separate vocals from instruments.
Each selected song is split into consecutive chunks across the song—not just its first few seconds. For example, a 3-minute song produces 18 ten-second clips. Any shorter leftover ending is excluded. Each training step learns from one clip, cycling through a shuffled pool; enough steps are needed to visit every clip. Longer clips provide more context but need more VRAM. This does not set the length of songs you generate or alter your original files.
ⓘ Shared style and clip length
Describe the shared sound: instruments, genre, vocal delivery, and production. Every selected song uses the same description. Clip estimates count complete fixed-length clips; shorter tails are excluded. Longer clips will require more training memory.
Choose a folder to inspect its audio. No matching text files are needed.
Review your dataset
This description is saved once in your training setup and applies to every selected song.
Each save creates a new project version. This prepares your setup; it does not start training or reserve the GPU.
Saved training projects
Train a style LoRA
Uses the saved project selected above, not unsaved form edits. Training joins Studio's GPU queue after existing jobs. Other apps are not managed by this queue.
ⓘ How much training?
Defaults to 2,000 updates, not 2,000 passes through every song. Each step learns from one clip. Use 20 steps only for a short technical smoke test. More steps can strengthen the effect, but excessive repetition can overfit and damage musical variety. Loss alone cannot tell you whether it sounds better.
ⓘ How strongly does it learn?
The peak update size; training warms up then tapers it. Start at 0.0001. Larger values can learn faster but become unstable or spoil the sound; smaller values may need more steps. There is no universal safe cutoff. Non-finite loss or gradients stop the run.
ⓘ How much can it remember?
Rank controls adapter capacity and memory use. Start at 8. Higher ranks produce larger files and can overfit a small collection. This version trains acoustic attention only; it does not retrain the song composer or guarantee singer identity.
ⓘ Checkpoints and stopping
Defaults to saving an intermediate adapter every 500 steps to audition later. Frequent saves use more disk space. Cancel waits for a safe boundary and saves completed training updates; closing Studio may force-stop after 10 seconds. Existing checkpoints survive. These adapter files are for inference, not exact optimizer-state resume.
ⓘ Models, caching, and the first run
Uses the local YuE2 and VAE folders from Models & runtime, independently of the inference backend. Requires a BF16 NVIDIA GPU; actual training memory use is still experimental. All selected audio clips are encoded first, even for a 20-step test. The cache is reused when source audio, VAE identity, and clip settings match. Original songs and model weights are never overwritten.
Start with the 20-step smoke test. No GPU training starts until you queue it.
Training runs
Train your artist sound.
Use full recordings, matching lyrics, and a shared style to train an Artist LoRA.
Artist Trainer · Experimental
Prepare Artist Trainer
These actions use the shared queue. Checks hash local files without GPU allocation. Downloads and dependency installation require confirmation; neither starts training. Setup remains experimental.
Encoder/companion terms · Reference pack · YuE2 · VAE · MERT
Use Check setup; results and errors appear in the Library run. Preparing paths records them only after verification.
Train the song-generation branch using whole recordings, full lyrics, and shared style. Separate from Style trainer's acoustic workflow. Experimental: check the verification status above; passing a short test does not guarantee voice similarity or a stable long run. Completed Artist runs appear in the Style / Artist LoRA selector. Select No score and Torch explicitly before generation.
ⓘ Trigger phrase
A distinctive phrase to associate with your artist. Future preparation will combine it with the shared style. Saving this setup does not train it yet.
ⓘ Lyrics are required for this new workflow
Choose the naming convention used by your folder. Each audio filename must have an exact matching text filename, without its audio extension. UTF-8 text only, up to 64 KB per song. Plain .txt files are treated as lyrics only when you explicitly choose that option; style-only files are not lyrics. Include complete words with section tags, then check each preview against the recording. We do not rename or overwrite source files.
ⓘ Shared style and full songs
Keep lyrics separate from this description. Uses complete 30–360 second songs, not the acoustic trainer’s short clips. No model quality, encoder compatibility, or voice identity has been verified by this scan.
Scanning and saving do not start a GPU job.
Saved artist setups
Train an Artist LoRA
Uses the saved setup selected above, not unsaved edits. At least two 30–360 second songs; the last selected song is held out for validation. Whole songs are used without silent clipping. Reference songs make up approximately half of training examples.
ⓘ Steps
Optimizer updates, with two song examples per update. There is no artificial maximum, but longer runs take proportionally longer and can overfit. Try 20 only for a technical smoke test; it is not a quality test.
ⓘ Rank
Adapter capacity. 64 follows the community recipe; lower ranks use less training memory but may learn less detail.
ⓘ Learning rate
The strongest update size reached after warmup. Higher values learn faster but can overfit or damage quality. The schedule below controls how quickly it becomes gentler.
ⓘ Learning-rate schedule
Auto warms up, then gradually reduces the rate to 20% at the final requested step. This supports longer runs but cannot prevent overfitting. Legacy reproduces the original curve, which tapers as if every run lasts 3,000 steps.
ⓘ Checkpoints
Keeps intermediate adapters so you can compare training lengths later. More frequent saves use more disk space. These are adapter snapshots, not exact optimizer-resume files.
ⓘ Timing methods
MMS forces every English lyric word onto the separated vocals. Whisper-assisted uses fine-tuned Demucs and Whisper large-v3 to recognize words, match them to your lyric file, and estimate missing positions. Only recognized word timing contributes to training; the full lyric text still conditions every song. A timing JSON is saved for review in the Library run. Whisper large-v3 weights download on first use.
ⓘ Lyric alignment weight
0.08 follows the community recipe. 0 skips all timing preparation and timing loss, regardless of the selected method; full lyrics still condition training. Higher values can overemphasize imperfect timestamps. MMS uses every aligned word; Whisper-assisted excludes estimated timestamps from the timing loss.
Uses Studio's shared queue. Progress, logs, cancellation, and saved checkpoints appear in Library → Open run. Cancellation waits for a safe boundary; separation/alignment of a full song can take time.
Make something worth hearing.
Your words, your direction. A whole new sound.
01 Source melody
SheetSage2 → YuE2Review the score below before creating your cover. This preserves a symbolic melody; it does not clone the source singer. Enter the lyrics separately.
Section tags guide the arrangement. Keep instruments and performance directions in Style.
Melody & score ABCOptional
Supply a native two-voice ABC score to use an existing melody. A supplied score bypasses the planner. Edit a copy; original run artifacts stay intact.
Prepare melody removes chord symbols and checks that the retained notes, timing and meter stay unchanged. Selecting one voice silences the other.
Checking local adapters…
ⓘ What does strength do?
0 has no effect; 1 is the starting point. Above 1 may distort audio or reduce coherence.
Trigger phrase: No phrase recorded
Added to the submitted style without changing your editor. Your choice is saved with each song.
Add a local LoRA / folder location
YuE2 acoustic Style adapters only. Adapters for other models cannot be used here.
Experimental GGUF: install audio.cpp separately, then connect it and download model files through Set up GGUF / Models. PyTorch runtime controls do not apply. Full songs and supplied-score covers are supported; Plan only and generated score export require torch.
Your sounds, all in one place.
Listen, revisit a score, or take a song in a new direction.
Your first song belongs here.
Create a song or transcribe a recording.
Its audio, score, settings, and progress will appear here.