README.md

metadata

title: MiniMax Music 3 Latent Refiner
emoji: 🎵
colorFrom: purple
colorTo: gray
sdk: gradio
sdk_version: 6.26.0
app_file: app.py
short_description: Restore damaged music in MiniMax Music 3 DAV latent space
python_version: '3.12'
startup_duration_timeout: 30m

MiniMax Music 3 — Latent Refiner v0.10

Restore damaged music while retaining the performance, timing, vocals, and arrangement. This Space runs the MiniMax Music 3 latent refiner — a 137M-parameter bridge transformer operating directly in MiniMax Music 3's continuous DAV latent space, conditioned on MERT frame-aligned features (layer 7), pooled CLAP music embeddings, and an SR3-style degraded latent stream with in-context reference tokens.

How it works

The refiner integrates a deterministic 32-step Euler bridge sampler from the damaged latent endpoint (t=1) back to the clean one (t=0), then decodes the refined latents through the DAV decoder. Inputs are resampled to 44.1 kHz stereo; clips longer than 30 s are processed in overlapping 30-second windows with a 2-second latent cross-fade (the training configuration).

Usage

1. Restore — upload damaged/degraded music (bandwidth-limited, noisy, quantized, clipped recordings) and press Restore. 32 bridge steps, 30 s windows by default.
2. Damage your own audio — apply the training-style degradation chain (bandwidth reduction, additive noise, bit-depth reduction, soft clipping) to a clean clip, then restore it in the Restore tab and compare against the original.

The bundled examples are public-domain instrumental music (FreePD/CC0, via mrfakename/cc0-music-captioned) with the degradation chain applied; the undamaged originals are included for side-by-side listening.

Notes

• FP32 is the verified release precision for the refiner.
• Inputs are trimmed to 120 s to keep ZeroGPU durations bounded.
• The minimax_music3_latent_refiner inference package is vendored in src/ (Apache-2.0, from the model repository).