README.md

metadata

title: Stem Restoration + Generation
emoji: 🎛️
colorFrom: gray
colorTo: yellow
sdk: gradio
sdk_version: 6.19.0
python_version: '3.10'
app_file: app.py
pinned: false

Stem Restoration + Generation

Upload a muffled or lo-fi song and get a cleaned-up version back. Under the hood we split it into instruments, then:

• Restorer — cleans up each muffled/damaged instrument so it sounds clear again (it fixes what's there; it doesn't add new parts). The glue model adds a small residual-flow detail head that samples back fine high-frequency detail; the plain deterministic backbone (advramp) is the selectable baseline and the default here.
• Generator — invents a missing instrument (mainly bass) that fits the song, like an AI session musician. Note: it fills a low end even for songs that may not have had a bass instrument.
• Glue stage (training-time) — restorer + generator are co-trained so their summed mix matches a real clean mix, judged by a multi-scale latent discriminator with feature-matching. It's a coherence objective; in practice it lands close to the deterministic baseline (both are selectable for A/B).

Everything runs in the compact SAME-L neural-audio latent space, on CPU.

Active models — restorers: ['remix_glue_v1', 'restorer_attn_advramp'] · generators: ['remix_glue_v1_gen', 'gen_advramp_v1']