yue2-mothersuperior-realaudio-tokenizer-v4: Tokenize, LoRA Train, and Generate Covers with YuE2

Provides NAR LoRA for the encoder and decoder to convert audio into semantic tokens, a feature not offered by the YuE2-3B model. Tokenize actual recording files directly to train artist-specific LoRAs, completing a workflow for generating new songs or covers.

Starting from version v5, a loss function that compares against actual audio was introduced to reduce the acoustic distance from the original. Full training and inference are possible in a 24GB GPU environment, and weights for ComfyUI are included to facilitate easy workflow integration.

Solves the real audio processing problem that was a limitation of music generation models. Useful for precisely replicating an individual artist's style or reinterpreting existing songs in new styles when combined with the existing YuE2 pipeline.