Skip to content

YuE2 Studio 5.0.0 Beta — ACE-Step and MiniMax libraries, engine updates, voice conversion

Latest

Choose a tag to compare

@timoncool timoncool released this 21 Nov 13:29
bfd35a0

Added

  • Bring songs from ACE-Step Studio and MiniMax Music3 Studio into the YuE2 library (#291, #525): open Library → Import → ACE-Step or MiniMax library. Select the old installation, or point Studio at the folder containing library.sqlite for a portable copy. The preview shows the songs and playlists before you copy anything. Your playlist order and song settings carry over, and the source library stays in place. You can import just the library records or copy the media into YuE2. Repeating an import skips records it already imported. Start a new take from a supported imported song with its native engine, including ACE-Step cover and repaint when the source latent is present.
  • Choose the engine, model and its controls in Create (#286, #389, #414, #459): the selector covers local YuE2, native ACE-Step and MiniMax Music3, audio.cpp, HOT-Step and OpenRouter. Create shows what you requested and what the service is using. OpenRouter lists music-capable models and exposes the parameters each provider advertises. It can generate without a local YuE2 model after you connect your account.
  • Install and update signed music engines from one page (#533, #537, #543, #548, #552, #561): on Windows, open Engines and Updates and choose Install beside ACE-Step or MiniMax. The first installation of the 5.0 runtimes downloads about 535 MiB per engine, with model weights listed separately. The page shows the available and installed versions, source, compatibility, model readiness and download progress. An engine in use finishes its job before a verified replacement takes over. You can cancel a download or restore the previous verified package, including offline. An update notice links to the page and is remembered across restarts for the same version.
  • Use LoRA and the original controls with native ACE-Step and MiniMax (#477, #480): ACE-Step lists compatible 2B and XL adapters. MiniMax exposes separate language-model and diffusion strengths. Replay restores the saved controls and adapter settings when the required files are available.
  • Generate and replay HeartMuLa songs through audio.cpp (#324, #332): Create takes lyrics and style tags with HeartMuLa's own sampling controls. Studio can install the weights from its model card. Replaying a saved take restores its settings.
  • Generate singing from text with Vevo2 (#324, #429): supply a voice reference and optionally a separate melody reference. The supported Q8_0 GGUF is 3,242,129,792 bytes. Select the folder containing the verified file in Create.
  • Generate instrumentals with Magenta RealTime 2 (#328, #440): describe a style and optionally supply an audio reference. The installation card shows where to place the supported Small F16 file (1.42 GiB).
  • Generate a sound scene with MiDashengLM-Gen (#324, #423, #440): fill in the music, speech, effects and ambience you want. Speech is optional. The duration field is a budget, and the resulting clip can be shorter. The installation card identifies the Q8_0 file (3.15 GiB).
  • Generate and replay with experimental MiniMax-H3 (#326, #418): the local audio.cpp option takes a prompt and stores the effective controls with the take. The model card provides manual weight setup and checks the required files. This is a separate option from native MiniMax Music3.
  • Generate with Stable Audio 3 Small Music through audio.cpp (#338): choose Small Music in Create and set the duration, seed and diffusion controls for that model. The Q8_0 route is separate from HOT-Step Medium.
  • Generate and train HeartMuLa or Stable Audio 3 Medium through HOT-Step (#362, #369): Create and MCP use the appropriate model controls. Train an adapter from the common training workspace and select the resulting GGUF in Create on its matching model. Medium uses its own HOT-Step package. Stable Audio 3 Small Music remains a separate audio.cpp option.
  • Create reference-guided covers with MuLaCover (#334, #342, #400, #439): choose the song to cover and a reference recording, then select the audio.cpp or HOT-Step route. The install card shows where to place the MuLaCover Q8_0 weights (7.66 GiB) and reports the resampler separately.
  • Synthesize or convert singing with SoulX-Singer (#406): select the mode in Create and provide the required score or reference audio. The result is saved to the library with its source and settings.
  • Create sound effects from a description or video (#442, #444): MOSS-SoundEffect takes a text prompt. ControlFoley creates audio for a selected video and leaves the source video available. Both save their results in the library.
  • Replace a sung voice and compare it with the original (#322, #351, #364, #394, #450, #550): find an RVC voice by filters and listen to a preview where the catalogue supplies one. Add your own supported v1 or v2 voice with an optional retrieval index. With Seed-VC, choose a reference recording instead. The new take keeps the accompaniment and goes into the library for A/B comparison. Speech clarity varies with the selected voice and recording.
  • Train a voice from your recordings on a compatible Windows NVIDIA system (#391, #408, #421): install the native Burn trainer from Studio, prepare a dataset, watch progress and apply the resulting voice to a song. Cancelling preserves completed checkpoints. Existing RVC voices in .pth/.index format still work.
  • Choose the YuE2 base for LoRA training and audition its checkpoints (#312, #453, #454, #460): select a supported base and keep its identity, including the decoder companion, with the adapter. Studio checks the match before applying the adapter. Dataset preparation can keep existing lyrics and skip speech recognition for the other songs. The audition playlist renders the same score and seeds against the base and saved checkpoints. A compatible Windows trainer can use CUDA 12.
  • Play a score several ways while keeping its notes and seeds (#451): choose how freely YuE2 writes the score and how it performs it. Create variations queues other readings of the score you edited. The numerical settings remain editable.
  • Find the key and chords of a backing track (#549): analyze a selected recording in Create, review the suggested harmony, and choose whether to use it. The same analysis is available through MCP.
  • Read, play and print a songbook from the score (#307, #314, #327, #509): transpose the chords, set a capo, use numbered notation and print an A4 lead sheet with lyrics and guitar or ukulele fingerings.
  • Organize the library into albums with artwork and track order (#337): create an album, edit its cover and arrange its songs. The same order and artwork are available through the library tools in MCP.
  • Mix and record in the DJ workspace (#353, #517, #538, #558): use two decks, loop and scratch within a track, mix stems where available and save a performance. You can discard a pending recording or retry saving it.
  • Compare a selected region in Edit Track (#467): listen to the original and edited audio over the same region and open the selected take in the library.
  • Install an optional YuE2 HIP runtime for supported Radeon adapters on Windows (#447): Studio checks the adapter and offers the corresponding package. The performance tables give the measured HIP and Vulkan results on the same Radeon iGPU and separate cold and warm timings for NVIDIA. The GPU benchmark table records the hardware, models and settings for each result.
  • Save timed lyrics as an LRC file (#431, #521): open a song's details after lyric alignment and choose Download LRC beside SRT. The Windows Save As dialog writes the file to the selected location. Reported by jiangqiang1996.

Changed

  • Reuse and repair verified files from another Studio installation (#283, #475, #546): choose an existing library to check its models and components before downloading. Files that match the expected size and SHA-256 can be reused. A damaged managed file can be repaired, optional downloads have their own progress and cancel controls, and verified installed packages can be reused offline. The public signed channel supplies component updates and mirrors.
  • Install supported tools from their original ZIP when a download fails (#241, #278): the error keeps the failed download details and shows the source link and exact location for the archive. Place the current ZIP there and repeat the same install action to verify and extract it. This includes the karaoke runtime, Captioner and HOT-Step training package. Reported by ryan20261015.
  • Use the newer native speech recognition engine for karaoke and lyric alignment (#481): the transcribe.cpp component is pinned to the tested 0.4.2-based build, with packages selected for the platform and backend.
  • See what the current audio.cpp task is doing (#541, #559): the job card reports stages around model execution and audio preparation. A cancelled task keeps its terminal status.
  • Keep audio export available when Ableton cannot use the score (#530): the exported project still contains the song and its available stems when the score comes from another engine. The export explains when it omits the score tracks.
  • Keep a usable score when generation reaches its limit (#529, #534): a score cut short by its length or memory budget can be kept and continued. Create explains when the requested duration is a poor fit for the score settings.
  • Control Windows song notifications in Settings (#524): the desktop remembers your preference between sessions.

Fixed

  • Re-rendered songs stay in the library when you delete or restore the original (#523): in 4.1.1, deleting a short original could also move the longer take you had just made to the trash. Delete and Restore now act on the selected song. The +60-second take keeps its own library entry and audio. Deleting or restoring the original changes only the original. Reported by Tapatvde.
  • Done remembers the installed model set you selected (#279): changing to another installed set in Settings saves the selection before closing. It is still selected after a restart. Reported by IAMJOYBO.
  • ACE-Step can read LoRA adapters from long Windows paths (#528): the Windows runtime accepts the long and Unicode paths produced by Studio when discovering adapter files.
  • The chosen training recipe and seed survive a trip back to the dataset (#277): manual and automatic start both use your selection. On Windows the checkpoint writer also keeps the previous complete checkpoint when interrupted during a write in a long path.
  • A Cover with blank lyrics keeps the instrumental instruction (#234, #434): both composer modes pass the explicit instrumental cue when the lyric field is empty. Reported by alexhanguery.
  • Reusing a song keeps its model, instrumental mode and adapter settings (#535, #551): Reuse Prompt restores those choices before the next request, and the lyric view follows the selected take.
  • A user-selected GPU is remembered for later launches (#536): Studio keeps the saved device identity when it reads older settings and uses it to resolve the selected card.
  • Completed scores can still be read after a Studio restart (#293, #368): the service saves completed score results so they remain available under their original job IDs. Engine job errors include context that helps identify the affected request.
  • Stopping the Linux service respects active component downloads (#513): an active download includes its first network check. Cancelling the transfer and then stopping an idle service still works.
  • An immediate retry shows the current download state (#519): retrying native voice-trainer installation starts a fresh progress record before the interface resumes polling.