YuE2 Studio 3.7.0

Added

• Cleaner vocals with Mel-Band RoFormer. Settings - Models - Stems now offers Mel-Band RoFormer voc_fv6 (Q8_0), Dub Studio's separator run by BSRoformer.cpp, beside HT-Demucs: it splits a track into the vocals and the instrumental, more cleanly, and the vocals of vocals-only songs and of training are separated with it too; on an NVIDIA card it runs on the engine's cuBLAS, elsewhere on the processor. HT-Demucs stays the default for six stems (#108).
• Sing again with new words. A song's menu - Sing again with new words: the song comes to the create form with its score and both seeds, without the take that sang the old words, so the new lyrics are what is sung; under the lyrics the studio names every sung line whose syllable count left the old one, since such a line moves the melody. Agents count syllables with lyrics_syllables (#108).
• LoRA from your own folder, without a copy. LoRA page - Add from folder: every LoRA of a folder you keep for all your programs (ComfyUI, a trainer) and its subfolders is taken in place - a folder with adapter_config.json is one LoRA, elsewhere each .safetensors is one - by a hard link, copied only from another drive, and the studio says which; agents use lora_import_folder (#98).
• Instrumental only. A switch on the create page above Create: the voice of the score turns into rests before anything is sung (a score is written first when the form has none), the lyrics keep only their section tags, and the Instrumental LoRA is added to the song's LoRA list, downloaded first when it is not installed; the button then reads Create instrumental. Agents set it with instrumental in create_form_set (#115).
• A saved stem names its song's tempo and key. A vocal or any other stem saved for a DAW is named like Song - vocals - 109 BPM, D major.wav, read from its score (YuE2 Studio #108).
• A unique pattern for every song. Settings - Covers - Pattern - Unique pattern of each song: the shapes come from the song's lyrics (a song without lyrics, its title) and the colours from its style, out of the pattern style's own palette, so songs of one style share their colours and new lyrics draw new shapes; a changed style or text draws the pattern again. The choices now read Pattern, Photo, Generation; a photo for the style stays the default (ACE-Step Studio #52).
• Your own agent CLI writes the words. Settings - Models - Writing assistant - Agent (CLI): Claude Code, Codex, opencode, Antigravity or Muse Code installed on this computer writes the lyrics, styles and layouts on your own account with it, one hidden run per draft, without tools; the panel shows which are installed, how to install and sign in, and tests one with a short draft. Antigravity runs only sandboxed. Agents see the CLIs with assistant_cli_agents (#112, thanks to GuilhermeAmado).
• An agent reaches everything the studio does. A test now holds every route of the studio to a tool or to a written reason, and the agent gains what it lacked: the network proxy and the HF-Mirror (get, set, test), access from the network, the Activity journal, the look of tracks without a cover, photo and pattern covers and photo search, video background search, a song's note, its own MIDI file, a playlist by itself, the score's MIDI export and import, instrumental, vocal octave and section marks, the graphics cards, the downloadable parts' state, and taking an ended job's card off the page.
• Lyrics in a video clip, found where they are. The video screen's Text tab always shows the lyrics overlay: a song without karaoke timing gets a button that makes it there, and the overlay turns on when it is done; without karaoke set up it says where to set it up. Until now the overlay section was hidden for such a song (#118).
• Fade in and fade out. The create page's extra settings take seconds for a fade at the start and at the end of the made song, so an ending no longer stops abruptly; the studio applies them to the song it keeps, Reuse Prompt and a re-render keep them, and agents give fade_in and fade_out (#109).

Fixed

• Karaoke on an instrumental no longer hangs on "downloading": the studio checked the song after fetching the recogniser and left that download's line running when it then refused the track; the song is checked first, and a finished download ends its line.
• LoRA training on Linux (#126 by boiler): the training page read the engine's CUDA build, which the Linux package never has, and refused a CUDA 13 card; it now reads what the card and driver run, and on Linux the trainer you build (YUE_TRAIN_BIN, YUE_CAPTION_BIN) links the system's CUDA. Karaoke on Linux takes the CUDA build of transcribe.cpp the same way.
• A long step no longer gets the engine restarted (YuE2 Studio #129): an engine inside a long step does not answer its health check in time, and the studio took that for a crash and restarted it, losing the song; while a job is running and the engine process is alive, it is at work.
• Auto stays on the graphics card after a silent engine exit (YuE2 Studio #127, #128 by ramishi): an exit with no device error in its log (out of memory, a kill, a slow start) no longer moves Auto to the processor for the rest of the session.
• The create form keeps its words (ACE-Step Studio #54): the name, the style and the lyrics stay across restarts, as the length and the file format do, until Reset clears them.
• The training page says where to install the card runtime. When preparation measures tempo on the processor, its notice now names the way to the graphics-card ONNX Runtime for this machine's card: Settings → Models → Optional extras → Split into stems → Graphics card on NVIDIA, Karaoke timings → Graphics card on AMD and Intel (DirectML).
• A computer's memory as its box says it. The studio counted RAM in 10⁹-byte gigabytes, so a 32 GB computer read 34 GB (in the issue form too); memory is now counted in 2³⁰-byte gigabytes, as video memory already was, and so are the model sets' needs.
• A cover from a FLAC, OGG or M4A recording. The engine reads WAV and MP3 only, and any other recording stopped at "cannot decode audio"; the studio now decodes it itself and gives the engine a WAV (#121, #93).

Changed

• Karaoke's Parakeet on transcribe.cpp. Parakeet v3 and Parakeet Ultra run as GGUF Q8_0 on transcribe.cpp, the engine Dub Studio and Speech Studio use, instead of ONNX Runtime: its CUDA build on an NVIDIA card with a CUDA 13 driver, its Vulkan build on any other card (and on Linux without the system's CUDA 13 cuBLAS), its processor build without a card; on a Mac Metal on Apple silicon and the processor on Intel. The files have Dub Studio's names and folders, so one program's copy serves the others, and the model leaves the graphics card when a recognition ends, so the card is the music engine's again. A saved fp32 choice becomes Parakeet v3 (full weights are not kept), and the ONNX Parakeet folders of 3.6 and earlier are removed at start.
• Whisper on the graphics card keeps int8 weights (int8_float16) instead of float16.
• Audio to MIDI and the training pack are one download. The MIDI transcriber, the trainer and the captioner come from one HOT-Step build that shares its graphics-card libraries, in one folder: 133 MB once instead of 126 and 132 MB apart, and installing either one installs both (YuE2 Studio #104).

Also in this release (prepared as 3.6.1, not published on its own)

Fixed

• An agent sees the window as you do. Through the studio's MCP it now sees folded sections (open or closed) and the options of open lists, reads a field by its own label rather than by the options inside it, and a field it fills while the window is in the background saves as after your own click.
• LoRA preparation says which song does not fit and why. A song longer than the model's 24576 tokens (prompt, score and audio together) passed preparation and failed at its end with "prepared manifest failed reader roundtrip"; the trainer now names the song and its sizes, and a roundtrip failure carries the reader's reason (#110).
• Karaoke and stems on a Mac. The Mac build downloaded the Windows ONNX Runtime and could not load it; it now takes Microsoft's macOS arm64 build, runs Parakeet and the separator on the processor, and the app may load that library (it is signed by Microsoft, not with the app). Whisper and the CUDA and DirectML parts are not offered on a Mac, where they cannot run (#105).
• A download Hugging Face does not answer says why and what to do. The error kept only "error sending request" and lost the cause (the name not found, a timeout, a refused connection); it now names it, and when the model comes from Hugging Face it points to Settings - Network, which can take models from the HF-Mirror instead.
• Karaoke that cannot start says why and ends. An ONNX Runtime the studio could not load stopped the recogniser with a panic: the karaoke stayed "running" for ever and every later try answered "Mutex poisoned" until a restart. The library is now loaded before any work and a failure is that work's error, with the library's path; a recogniser that stops ends its karaoke as failed (#106).
• The song details button is back on wide windows. 3.6.0 hid it on the create page whenever the details could sit beside the list; it shows at every width again (#99).
• The Linux AppImage starts its engine again. The AppImage put the MIDI tool's older ggml library on the path every program inherits, so the engine loaded it instead of its own and stopped at once ("undefined symbol: ggml_prec_set_acc"). Every program the studio starts now loads the libraries beside it first: the engine, the writing assistant, the trainer, Audio to MIDI and the karaoke recogniser (#97). Thanks to angelside, who found the cause.
• The writing assistant works on Linux. It downloaded the Windows builds of llama.cpp and could never start; it now takes llama.cpp's Linux builds for the processor, Vulkan and CUDA (#114).