lyramakesmusic

lyramakesmusic / sa3-inpainter-ui

Public

SA3 medium audio inpainter — MLX SAME-L decoder + FastAPI + vanilla Svelte UI

10
2
85% credibility
Found May 23, 2026 at 10 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

This is a browser-based audio editing tool that lets you paint over sections of a sound file and regenerate just those parts using artificial intelligence, while keeping everything else exactly as it was. You can also generate entirely new audio from text descriptions or transform existing recordings into different styles using add-on plugins called LoRAs. The whole thing runs locally on your computer — your audio never leaves your machine. It's designed specifically to work smoothly on Apple Mac computers using custom MLX code.

How It Works

1
🎨 Open the audio painting canvas

You load your audio file into a beautiful browser canvas that shows colorful waveforms and spectrograms, with a timeline you can zoom and scroll through like a map.

2
🖌️ Paint the sections you want to change

You click and drag across any part of the audio to mark it for regeneration — think of it like painting with an eraser over the parts you want to remake.

3
💬 Describe what you want to hear

You type a short phrase like 'warm acoustic guitar with gentle percussion' to tell the AI what style or mood you're aiming for.

4
Choose how you want to transform your audio
🎯
Inpaint (keep the rest)

Only the painted regions change — your original audio stays perfectly intact.

🎭
Transform (change everything)

The entire audio gets reimagined in your described style.

5
⚡ Watch your audio get created

A progress spinner shows while the AI thinks, and as each section finishes you see the spectrogram fill in like a photograph developing.

🎵 Play back and save your creation

Your edited audio plays back with a smooth playhead that you can click anywhere to scrub through. When you're happy, you download the final file.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 10 to 10 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is sa3-inpainter-ui?

This is a browser-based audio workstation for Stable Audio 3 medium that lets you paint regions on a spectrogram and regenerate just those sections while preserving everything else. It combines a Python backend with a vanilla Svelte frontend, running the heavy decoder work on Apple Silicon via MLX while using Torch for the diffusion model on MPS. You get three generation modes: text-to-audio from scratch, audio-to-audio variation, and true inpainting where you selectively regenerate masked portions of an uploaded sample.

Why is it gaining traction?

The killer feature is native Apple Silicon support without flash attention -- the author built an MLX port of the SAME-L autoencoder decoder that fixes broken sliding-window attention on Mac. The UI is unusually polished for a research project: scroll-to-zoom anchored at your cursor, shift-scroll to pan, click-to-scrub the playhead, and a lowpass filter that ducks masked regions during playback so you can hear what the inpainted sections sound like in context. LoRA stacking with per-strength sliders opens up style transfer without leaving the interface.

Who should use this?

Musicians and sound designers who want surgical control over AI-generated audio without cloud dependencies. Researchers benchmarking inpainting quality on Apple hardware. Developers building audio editing tools who want a reference implementation of the SA3 medium pipeline with working MLX support. Not for production deployment yet -- the 10 stars and hardcoded paths suggest early-stage work.

Verdict

A technically impressive playground with a genuinely useful feature set, but treat it as a development tool rather than a finished product. The 0.85% credibility score reflects a small but active single-developer project with solid architecture documentation and end-to-end parity tests. Clone it if you're on Apple Silicon and want to experiment with SA3 inpainting locally; wait for more stars if you need battle-tested tooling.

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.