dada-bots

dada-bots / underfit

Public

Stable Audio LoRA Trainer of salty goodness

audio-ai lora lora-fine-tuning music-ai stability-ai
19
2
85% credibility
Found May 28, 2026 at 19 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

Underfit is a web-based dashboard that helps you create custom LoRA adapters for Stable Audio 3 — a music-generating AI model. You load your own audio files, the tool encodes them and trains a small adapter on your GPU, and you get a downloadable file that lets you generate new music in your style. The dashboard shows live progress with audio previews and loss charts, supports multiple training runs at once, and lets you blend adapters together for creative results.

How It Works

1
🎵 You gather your music

You collect audio files of music you love — your own tracks, a favorite artist's style, or sound effects you want to recreate.

2
🎛️ You open the dashboard

You launch a web dashboard in your browser that shows you everything at once — your datasets, training runs, and generated audio previews.

3
📁 You create a dataset

You point the dashboard at your audio folder, and it automatically reads any tags or titles from your files to understand what each track is about.

4
✨ You train your style

You click a few buttons to start training — the dashboard encodes your audio, trains a small adapter on your GPU, and generates sample clips every few minutes so you can hear how it's learning.

5
You listen and decide
🔄
Keep training longer

If you want more creative variation, you let it train further — the model keeps generating fresh demos as it learns.

💾
Save a checkpoint

Once it sounds the way you want, you download the trained adapter file and use it to generate new music in your style.

🎶 You make music in your style

You use your trained adapter to generate brand-new tracks, blend it with other styles, or apply your sound to existing audio — all from the same dashboard.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 19 to 19 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 underfit?

Underfit is a web-based dashboard for training LoRA (Low-Rank Adaptation) adapters on Stable Audio 3 models. In plain terms, it lets you create custom audio generation models by fine-tuning on your own music or sound effects. You upload audio files, configure training parameters through a browser interface, and the system handles the heavy lifting—encoding your audio into latent representations, running the training loop across GPUs, and generating sample clips as it learns your style. The project is written in Python and builds on PyTorch, with support for NVIDIA GPUs and a Colab notebook option for users without local hardware.

Why is it gaining traction?

Most LoRA training tools are command-line only, which makes experimentation slow and intimidating. Underfit flips this with a real-time dashboard showing loss curves, GPU memory usage, and auto-generated audio demos every few hundred steps—so you can hear your model improving as it trains. It supports multiple LoRA variants (including DoRA and BoRA), lets you blend multiple adapters together for inference, and includes audio2audio style transfer and inpainting features. The setup script handles dependency management automatically, making the installation process far less painful than comparable tools in the stable audio ecosystem.

Who should use this?

Music producers who want a model that sounds like their style—not generic AI output. Sound designers building specialized SFX libraries. Developers experimenting with custom audio finetunes without needing to piece together separate encoding, training, and inference pipelines. If you're comfortable with a Linux GPU box or Colab, and you have 10+ minutes of cohesive audio to train on, this tool makes the process significantly more accessible than writing training scripts from scratch.

Verdict

At 19 stars and 0.85% credibility score, underfit is a niche tool built by someone with real domain expertise. The documentation is thorough, the feature set is well-thought-out, and the auto-resume and multi-GPU capabilities suggest a production mindset despite the early stage. Maturity is low—expect rough edges and limited community support. If you're serious about audio LoRA training, this is worth evaluating; just know you're adopting an early-stage project where the learning curve is still steep.

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.