drbaph/yue2-mothersuperior-realaudio-tokenizer-comfyui

YuE2 Mothersuperior real-audio tokenizer — ComfyUI safetensors

Safetensors conversion of Mothersuperior/yue2-mothersuperior-realaudio-tokenizer-v4 (pinned revision f2278a2e005dc4ecc421c53a0929f62b3aeb2280) for use with ComfyUI-YuE2-Trainer.

This is a format conversion only — no weights were retrained, merged, or modified. The original repo publishes pickle (.pt) checkpoints; these are the identical tensors saved in the safetensors format so ComfyUI can load them without ever invoking torch.load at runtime. Each file embeds provenance (source filename + SHA-256, format tag, license) in its safetensors metadata header.

Files

File	Source .pt	Size	Contents
tokenizer_head_joint_v4.safetensors	tokenizer_head_joint_v4.pt	163 MiB	MERT-v2-FullSong layer-20 features (instance-normalized, 25 Hz) → 32,768 YuE2 semantic codes. 8-layer transformer, d=512, 512-frame windows.
nar_lora_joint_v4.safetensors	nar_lora_joint_v4.pt	134 MiB	Rank-32 NAR LoRAs on nar_self_attn.{q,k,v,o}_proj + nar_mlp.{gate,up,down}_proj (28 layers) plus full vae2llm/llm2vae replacement weights and biases.
minted_regularizer_pack.safetensors	regularizer/minted_regularizer_pack.pt (from yue2-minted-corpus)	83 MiB	4,732 YuE2-generated songs as {name, src, style, lyrics, codec} records.
minted_regularizer_pack.jsonl	—	4 MiB	Companion metadata for the pack. Keep it next to the safetensors file.

Usage notes

• The NAR file is not a pure LoRA. Its metadata flags io_semantics: full replacement weights and biases, not LoRA deltas — the vae2llm/llm2vae projections must be applied as replacements, not scaled by a strength factor. ComfyUI-YuE2-Trainer's loader handles this correctly.
• Load via ComfyUI-YuE2-Trainer (category YuE2/Training/Artist Training (recommended)): the YuE2 Mothersuperior Assets node downloads or picks up these safetensors and manages the MERT model and cache for you. Only the AR artist-training path needs the head + regularizer pack; the NAR adapter is optional.
• The AR LoRA you train with these assets is a separate, standard ComfyUI LoRA — apply it with the stock LoraLoader node (both MODEL and CLIP), not LoraLoaderModelOnly, because the AR branch of YuE2 lives on the CLIP/text-encoder side.

Verifying the conversion

Each safetensors file records the SHA-256 of its source .pt in its metadata (source_sha256). To check a file:

from safetensors import safe_open
with safe_open("tokenizer_head_joint_v4.safetensors", framework="pt") as f:
    print(f.metadata())

Recorded source hashes (revision f2278a2):

• tokenizer_head_joint_v4.pt: d23c4f757a05f031134b8471ec84245ec2338966516e1a9e26a17ff300a5f87e
• nar_lora_joint_v4.pt: df175dbf9405a8e15b2c3f8dbdcc97303575f763787f227b03029020e28102fe

Provenance and terms

Weights derive from m-a-p/YuE2-3B and the Mothersuperior real-audio adaptation. CC BY-NC 4.0 — non-commercial use only. Full training recipes, measured held-out accuracy (16.1% top-1 on YuE2's own songs), and the upstream scripts live in the original model repo — please read its model card and credit the Mothersuperior authors. Converted files are provided as-is, with the same terms as the source.