YuE2 Studio 0.8.0

Added

• Radio filters. Search and combine multiple genres, languages, and vocal/instrumental choices in a compact filter dialog. Selections persist and apply to playback and shuffle within the selected source.

• Live Night / Chill selection. Additional templates and persisted template selection guide automatic generation and eligible catalog playback.

• Live jukebox fallback. Writing credits/quota or generation failures pause new requests while current audio and ready songs continue. Past library songs shuffle once per cycle; missing/corrupt files are skipped. Resume generation explicitly after resolving the problem. Live excludes raw errors and private panels from broadcast capture.

• MCP 1.1.0 adapters. Nineteen tools expose advanced YuE2 controls, planned ABC, Stable/Woosh effects, stems, lyric synchronization, cover art/transcription, audio export and typed Studio mixing with validated payloads and readiness checks.

• Shared song queue. Several people (for example the studio PC and a browser on the home network) can queue songs at the same time; they are made one after another on the GPU. The Create page and Library show the line with each song's owner, position (Up next, Waiting in line · N ahead of you) and live progress. Anyone can keep clicking Add to queue. Each browser picks a display name. Only the person who queued a song, or the studio PC, can cancel it. The library refreshes for everyone when any song finishes.

• Keys → Writing model dropdown. The Model box now lists the models the chosen provider actually offers (live from NVIDIA NIM, OpenAI, Groq, xAI, Anthropic, Gemini, or your Local LLM server), with song-friendly picks under Recommended and image/speech/embedding/safety models hidden. Type a model ID… still allows any ID; a warning shows when a typed ID isn't in the provider's list.

Security

• The local service now only accepts commands from Studio itself (the desktop window, pages served by Studio on this PC or the home-network address). Other websites open in a browser can no longer control Studio in the background, and DNS-rebinding hosts are refused.

Fixed

• Global EQ tuning. Gentler presets preserve deep bass while reducing muddiness. Playback gain accounts for the combined filter response to leave headroom across Radio, Library, and Live, including custom sliders.
• Writing responsiveness and logs. NVIDIA GPT-OSS uses low reasoning effort; compatible Nemotron requests disable thinking. Logs identify the actual provider, model, action, request ID, elapsed time, waiting state, completion, and failure without recording keys or prompts.
• Writing vault normalization and API updates now use provider-specific defaults, repair known inherited IDs, and migrate the retired NVIDIA MiniMax model to the current NVIDIA default. Custom model IDs remain supported.
• YuE2 no longer imports the inherited caption corpus. The compatibility statistics route reports it disabled; installer payloads exclude dormant Music3 workers and caption assets. Song generation remains exclusively YuE2.
• Credential audits cover all saved provider keys, with values withheld; private settings, old executable backups, and Windows shortcuts stay excluded from source publication.
• Sidecar API metadata now reports the current 0.8.0 application version.
• MCP server: song folders with spaces (most of the library) now work with update_song/rate_song; replies are always UTF-8 on Windows (non-English characters previously broke the reply for every MCP client); the connected agent's name (Claude, Codex, Grok) labels its activity and queued songs.
• NVIDIA NIM's free tier gets 150s per writing call instead of 90s (lyrics took 82–98s in testing), and the desktop bridge waits up to 480s. Timeouts now name the real provider instead of "Local LLM timed out".
• Job-starting endpoints return the song's queue position immediately.
• Switching the Writing provider no longer keeps the previous provider's model (e.g. gemini-3.6-flash under NVIDIA NIM).
• The default NVIDIA NIM model minimaxai/minimax-m3 was retired from NVIDIA's catalog; the default is now openai/gpt-oss-20b, tested against the live NVIDIA service.
• Finished jobs are trimmed (last 150 kept) so the job list no longer grows for the life of the process.

Windows: download the x64 setup executable below. Model weights and optional GPU runtimes are downloaded separately from Models. Existing users should close Studio before upgrading.

YuE2 Studio 0.6.1

YuE2 Studio 0.6.1

Full notes: CHANGELOG.md

Fixed

• SheetSage2 install no longer fails after rewriting config.json for the local MERT encoder.
• Library playback uses the native audio element. Web Audio was muting songs in the desktop WebView.
• The “Song ready” banner no longer keeps a red elapsed clock after generate.

Windows installer

Download YuE2.Studio_0.6.1_x64-setup.exe. Models are still installed from inside the app.

YuE2 Studio 0.6.0

YuE2 Studio 0.6.0

Full notes: CHANGELOG.md

Added

• Remix this song from the saved YuE2 score (keep melody, or keep melody and chords).
• Cover from audio with optional SheetSage2 transcription (install from Models; CC-BY-NC).
• Video Studio scene tiles (including chroma keys). Scene color is backdrop only.
• Visualizer presets: Analog, Equalizer, Cascade, Stereo, Silk, Dots; reworked Sunburst, Mesh, Peak.
• /api/video/prepare unloads YuE2 before canvas capture.

Changed

• Official YuE2 compact style/lyrics; Off-mode CFG 1.01; compact voice compile applied on generate.
• Peak/Bars/Wave/Grid/EQ anchored to the bottom in landscape and portrait.
• Faster Video Studio capture: no per-tick looping-video seek, cached note sprites, skip extra glow on photo/video backgrounds, cap draw rate during render, faster FFmpeg.

Fixed

• Library and Video Studio playback loops on status polls.
• Library GET no longer renames playing audio files.
• Particles draw with visualizers on every scene, including black.

Removed

• Weak/duplicate presets: Aurora, Skyline, Flow, Rings, Spiral, Helix, Vista, Contour, Disc, Ribbon.
• Primary/Secondary color pickers.

The Windows installer is not attached to this tag yet. Source on main is 0.6.0. Build with pwsh -File scripts/build-installer.ps1 if you need YuE2.Studio_0.6.0_x64-setup.exe.

YuE2 Studio 0.5.1

Full Changelog: v0.5.0...v0.5.1

YuE2 Studio 0.5.0

YuE2 Studio 0.5.0

Start here: Complete user guide. Covers setup, every main workspace, prompting, Studio editing, export, troubleshooting and backups. A copy is included in the installer and attached below.

• Added Sony Woosh-Flow as an optional GPU sound-effects engine, with negative prompting, CFG and sampler controls.
• Added Woosh to Models: pinned, verified downloads; free-space checks; progress and cancellation; tokenizer, decoder, text encoder and private Python 3.12/CUDA runtime setup. Installed features remain hidden.
• Added Combine sound tracks and Restore original tracks in Studio. Combining renders selected imported lanes with their timing, overlaps, levels, fades and effects.
• Removed the indefinite "Calculating ETA" placeholder; elapsed time stays visible when no estimate is available.
• Fixed mono sound playback and export so imported effects are audible in both ears.
• Included the recent Stable Audio sampling controls and optional Gemini prompt helper. Simple prompts are recommended for Woosh.

Windows installation

Download YuE2 Studio_0.5.0_x64-setup.exe below. The updated installer includes Python, FFmpeg, the current application and model setup code. Models are optional downloads from the Models page; no model weights are bundled. Close the running app before updating.

Woosh-Flow generates up to five seconds per clip. Its released weights are licensed CC-BY-NC (non-commercial).

Validation and privacy

• Windows installer build and payload privacy audit passed.
• 168 backend tests passed, including mono export, combination/restoration and Woosh installer tests.
• Existing local Woosh installation verified through the new installer workflow.
• Source excludes model weights, private runtimes, credentials and generated user data. The installer excludes API keys and personal data.

YuE2 Studio 0.4.0

YuE2 Studio 0.4.0 is a native Windows music studio for YuE2-3B.

• Windows installer creates the required folders and includes Python and FFmpeg.
• Models are downloaded only when selected in the app; already installed models are hidden.
• Includes song creation, library tools, Whisper lyric synchronization, cover art, stem separation, Studio and video tools.
• Includes a 22-screen tour captured from the Windows executable.

The source repository excludes credentials, personal outputs, model weights, installed runtimes, caches and build products. The installer payload was checked against the local Gemini key and contains no key vault or user library. Each user supplies their own optional cloud key.

Validation: clean source frontend and native Windows builds passed; the clean backend imports as version 0.4.0. Studio stem extraction completed in the app during the screenshot tour.

Download the x64 setup executable below to install, or use the source archive and follow README.md to build locally. The SHA-256 file verifies the installer download.