Mothersuperior Releases Real Audio Toolkit for YuE2-3B

September 15, 2026September 18, 2026

At a Glance

Values determined by this site’s code at collection time. Dates are JST.

Overview

Mothersuperior has released a real audio toolkit for the open music generation model “YuE2-3B", titled Mothersuperior/yue2-mothersuperior-realaudio-tokenizer-v4. It is a real audio tokenizer that accepts audio or music as input, encodes them into semantic tokens, and renders actual production-level latent representations through NAR branch LoRAs. This enables tokenizing one’s own recordings, fine-tuning artist LoRAs, and generating new songs or covers.

Specifications

• Base model: m-a-p/YuE2-3B
• License: cc-by-nc-4.0 (Non-commercial use only)
• Included files:
• tokenizer_head_joint_v4.pt: An 8-layer transformer (d=512, 512-frame window) that converts features from MERT-v2-FullSong layer 20 into 32,768 YuE2 semantic codes. Achieves 16.1% top-1 accuracy on YuE2’s own held data for songs – nar_lora_joint_v4.pt: Rank-32 LoRAs for 28 layers of nar_self_attn.{q,k,v,o}_proj and nar_mlp.{gate,up,down}_proj, as well as vae2llm/llm2vae

Performance and Quality

The model card of the original model publishes evaluation results on WildSongBench and SHS100K. Below are the evaluation results for full song generation on WildSongBench.

While the original YuE2 model records excellent scores in many metrics such as Musicality, SongBench Avg, MuLan, and AllMusicCaps, it falls short of other models in certain items like PER (Phoneme Error Rate). These are figures published directly by the authors.

Strengths and Use Cases

It is suited for tokenizing one’s own recorded audio, fine-tuning AR via LoRA using tracks from specific artists, and generating new songs or cover tracks. After being trained on 4,765 YuE2 self-generated songs, it has been adapted to real audio, and combining it with an additional regularization pack (Mothersuperior/yue2-minted-corpus) can prevent token grammar collapse with small artist sets. It is also possible to transcribe existing recordings with tools like SheetSage2 to obtain melodies and generate cover songs.

Differences from Similar Models

Unlike the original base models Music Generation Model “YuE2-3B" Released, Create Songs from Lyrics and Style and Repackaged Version of Music Generation Model “YuE2-3B" for ComfyUI Released, this model adds a real audio tokenizer and NAR branch LoRAs that are not officially bundled, making a specific difference by enabling direct tokenization using one’s own audio and artist-specific LoRA training.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 3.6B parameters (taken from the base model m-a-p/YuE2-3B)

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

How to Get It

• Repository: Mothersuperior/yue2-mothersuperior-realaudio-tokenizer-v4 (Hugging Face)
• Required additional data: Needs the regularization pack (regularizer/minted_regularizer_pack.pt) from Mothersuperior/yue2-minted-corpus
• Dependencies: Python 3.12 venv, yue2-infer (commit 92a73cc7), torch 2.10 + cu128, torchaudio 2.10, transformers, soundfile, scipy, safetensors, demucs

Related Articles

• YuE2-3B Music Generation Model Released for Lyrics & Styles
• Comfy-Org Releases YuE2 for ComfyUI: Open Music Generation

Sources

• https://huggingface.co/Mothersuperior/yue2-mothersuperior-realaudio-tokenizer-v4
• https://huggingface.co/m-a-p/YuE2-3B