YuE2 Studio 3.5.0

Update from 3.4.0.

Added

• Anonymous statistics and news from the hub. The start screen and Settings - Anonymous statistics have
  a checkbox, on by default, with which the studio sends once a day how many songs were made, failed or
  were cancelled, the model set used, its version, the OS and the class of the graphics card - never lyrics,
  prompts, audio or file names; What is sent shows the report, and DO_NOT_TRACK=1 or STUDIO_TELEMETRY=0
  turns it off entirely. News from the author arrive without an update, on top of the bundled ones.
• AB2 solver under Advanced (Riff's idea): one network evaluation a step instead of two, so the
  sound stage takes about half the time at the same steps; in our comparison it sounded the same
  as midpoint, which stays the default. Agents set solver.
• A song from a seed (Mothersuperior's hum-to-song, its first part): with "Continue the song from
  this score" the score is only an opening - hummed and transcribed, played on the keyboard or
  written - and the model writes the rest of the song on from it in its key and meter, the opening
  coming back as a hook; silent bars at its end are dropped. Agents set abc_continue.
• Chord bed beside Compose the score (gary4juce's idea): a score of chords over rests written at
  once in the tempo, key and meter you set, one section per lyrics tag, verses, choruses and a bridge
  each on a common progression of their own; the model writes the melody and the arrangement over
  it and the planning step is skipped. Agents use score_chord_bed.
• Chords and sections under Advanced, for the score the model writes (Yeufonic's idea): chord
  variety lowers the chords heard among the recent changes so the song stops looping one
  progression, a section can be kept from opening the way the one before did, chords outside
  the key can be favoured, and the score can be held to the lyrics' sections in their order.
  Off by default; agents set it with harmony on song_create and score_compose.
• Extend a song. Re-render takes an extension of 15 seconds to 2 minutes: the track keeps
  its own semantic tokens to the last frame and the model composes on from there.
• Words follow the score (HOT-Step's lyric schedule): with a supplied score each section's
  words stay out of the composer's sight until the score reaches them, so covers and long songs
  keep the voice with the band. On by default; a switch under Advanced.
• Realaudio decoder: Auto, On, Off. Auto decodes with the YuE2 checkpoint alone, as its
  authors, audio.cpp and ComfyUI render it, and merges Mothersuperior's decoder adapter only
  under a LoRA trained in the studio, which was trained over it. On and Off force it.
• Hugging Face mirror in Settings - Network (hf-mirror.com) for downloads where
  huggingface.co is slow or blocked, after wangsoft's fork.
• Score as PDF. The score view saves the notation as an A4 PDF, each system whole.
• The model set fits the computer's memory, not only the card's. On integrated graphics or the
  processor the weights live in RAM, so a laptop is now offered the set its memory holds, and every
  set that needs more memory than the computer has says so on the start screen instead of stalling
  half loaded. The studio no longer recommends the unquantised BF16 set: Q8_0 is near lossless at
  half the memory, and the BF16 set stays in the list.
• Match sections to the score for a cover (after HOT-Step's Cover Studio): the source recording is
  heard by the karaoke recogniser and each lyric block is tagged with the score section it is sung in.
  Words never change; a chorus written once is copied into later chorus sections, sections without
  voice get an empty tag, blocks not heard clearly keep their place and are listed. The proposal is
  shown before and after and applied only on Apply. Also an MCP tool, score_match_sections.
• Voice an octave down or up: the score view moves only the vocal line by an octave, the band,
  chords and key stay. The written register decides who sings, so an octave down brings in a man; and
  a vocal line whose middle leaves C4–A#5, where the model writes the voice itself, is flagged.
• Russian stress: a stress mark (U+0301, Alt+0769) after a vowel puts the sung stress on it; the
  lyrics hint says so.
• Engine progress on the card: the stage, its step counter and the time left.
• Video export settings: frame size from 1080p to 240p and AAC from 128 to 320 kbps.
• Songs up to 10 minutes. The window keeps its size, place and maximised state; the player its
  repeat and shuffle. Quitting while a song is made asks first and stops it; a song the studio
  was closed on starts again. From pull request #51 by
  stalexxx, which also runs the service on macOS with Metal.
• Instrumental from a score. A button in the score view moves every Vocal note to the
  instrument part with its pitch and length, as m-a-p's yue2-instrumental skill does, and keeps
  the chords; the instrument keeps what it played where the voice was silent.
• A note on every song, and its parameters sent to the form, shown as JSON and saved to a
  file, as in ACE-Step Studio.
• Parakeet Ultra int8 for karaoke: Moondream's fine-tune of Parakeet, quantized, a recogniser
  of its own beside v3, which stays as it was.
• Catalog: raspy rock-soul, quiet storm R&B and dark Sufi fusion LoRAs by becausereasons.
• The particle slider LoRAs of ntc-ai at their v2 release.

Fixed

• Long songs no longer fail now and then at the sound stage with a CUDA illegal memory access: the
  engine's attention keys are padded to the flash attention kernels' tile (the sound is unchanged,
  and that stage got faster).
• A cover's melody score no longer shrinks to a quarter of its size in the score view and its PDF:
  the voice's multi-bar rests are drawn bar by bar.
• A score the model wrote as garbage (no key, no meter, colons in the notes; LoRA strengths far past
  their limits do this) stops the song in seconds with the reason, instead of minutes of singing
  from it.
• A LoRA trained in the studio gets the style phrase it was trained on ("T, in the style of T.").
• Audio to MIDI no longer needs CUDA 13: it runs on CUDA where the engine does and on the
  processor elsewhere (Pascal, AMD, Intel); the new package is fetched once.
• The MIDI editor kept notes in place after an unknown meta event (files from Reaper); a file of
  instrument parts no longer turns one into the voice; Save MIDI of the score wrote a broken file.
• The exported video's visualizer moves as in the player; its background no longer freezes on
  zoom and pan.
• Models from a folder: the window says what was taken, copied and still missing, and the start
  screen shows why the engine did not start.
• Updating a LoRA from the catalogue left its old weights beside the new ones, and the engine
  refused the pair.
• Linux groundwork from SkySlider's fork: the graphics card is named on Linux, child processes
  end with the studio, and setup no longer claims a download that is not happening.

Changed

• A song is kept as the model made it. The engine hands over its float output and the
  studio encodes it once, changing nothing on the way: lossless 24-bit FLAC by default, written
  by libFLAC 1.5.0, the reference encoder; MP3 by LAME when chosen. Generation no longer
  normalises the peak: Normalise is a stage under Process, after mastering.
• Tags in every kept format (lofty): title, artist, genre, tempo, lyrics and cover go into
  a FLAC's Vorbis comments and picture block as into an MP3's ID3v2.4.
• Engine: yue2.cpp of 8 October. SheetSage2 reads MERT-v2 from a file of its own (F32 or
  Q8_0); the new transcriber is downloaded once in Settings - Models, and the studio runs without
  it until then. Covers transcribe exactly as before.
• Trainer: HOT-Step of 7 October, about 14% faster per step with identical weights.