SimpleTuner/open-rvq-encoder-minimax-music3
Updated • 2
A community experiment to distill a compatible residual vector quantisation encoder from the hidden states and audio latents produced by MiniMax Music
Note Contains ComfyUI custom node, Diffusers patches, and the full v1-v4 model collection.
Note The first experiment that reached ~76% replay accuracy
Note Wider model footprint for another percent at ~77% replay accuracy
Note Applying representation learning via MERT to slightly improve acoustic signal
Note Instead of independent heads, use autoregressive conditioning and reach 87% replay accuracy with the highest acoustic signal integrity