The main drawback of modern generative music is its mathematical sterility. Models like YuE2 output finished compositions, but to the ear, they often sound constrained, lacking natural dynamics and high frequencies. The open-source project Partscaster 1.0 (formerly yue-studio) offers a fundamentally different approach: instead of simply running generation, it disassembles the resulting material and rebuilds it from scratch using real samples and analog audio signal chain emulations.
Optimization: Switching to C++ and Running on 8 GB VRAM
The key engineering breakthrough of the 1.0 release is the complete departure from the heavyweight PyTorch pipeline (which required at least 11 GB of VRAM) in favor of the GGML-based yue2.cpp engine. Converting weights to the GGUF format radically reduced system requirements:
Low barrier to entry: Generation now runs stably on GPUs with 8 GB of memory (successfully tested on an RTX 4060);
Quantization flexibility: The Q8-quantized model takes around 8 GB, while in lightweight mode it can fit into 4.5 GB;
Speed boost: On an RTX 4070 Ti Super, rendering a 2:47 composition dropped from 91 to 50 seconds.
Revitalizing Arrangements: Drum Fills and Samples
Partscaster analyzes the generated audio and replaces synthetic drums with high-quality acoustic samples of kicks, snares, cymbals, and toms, taking velocity into account. Because neural networks rarely use fills between song sections, the project adds an algorithmic "drummer" that automatically places accents before transitions between verses and choruses. For synth fans, classic drum machines are included: Roland TR-808, TR-909, LinnDrum, CR-78, and Simmons.
The bassline is reworked similarly: track notes are translated into real bass parts.
Guitar Double-Tracking and NAM Analog Chains
Fully isolating the original electric guitar using demixers like RoFormer is still difficult. Therefore, the developer implemented a session guitarist mode: the tool identifies the harmonic grid and generates clean parts for semi-acoustic or nylon guitars. To eliminate the characteristic "monophonic narrowness" of neural networks, the part is recorded in two independent takes (double-tracking) with stereo panning.
To tweak timbre characteristics, a dedicated "Instruments" section is provided:
Audio is routed through an effect chain based on NAM (Neural Amp Modeler) — neural network captures of analog pedals, amp heads, and cabinets;
Response latency when tweaking virtual knobs is only about 0.25 seconds without stopping looped playback;
A dialed-in setup can be saved as a custom preset for the entire track.
Spectrum Restoration and Control via MCP
Spectral analysis revealed that the YuE2 model cuts off frequencies around 10–16 kHz and drops crash cymbals by 6–30 dB compared to standard studio recordings. Partscaster programmatically reconstructs the lost high-frequency range and balances the cymbal levels against FLAC studio references.
Final tuning relies on a library of 100 genre profiles across 34 musical styles. Furthermore, the project can be controlled via the MCP (Model Context Protocol), delegating the entire rebuilding, sample selection, and mastering workflow to an external AI agent.
Where to Download and How to Run
The project is completely open source. The source code is available in the GitHub repository, and interactive before/after audio demos are hosted on the official website.
An NVIDIA GPU is required. The installer (the download size for models and sample libraries is about 20 GB) is provided as a background worker for Windows (no WSL setup required) and an off-the-shelf ./install.sh script for Linux, with support for remote deployment via SSH.