FragmentaBeta · 2025–2026

Fragmenta

An all-in-one pipeline for training and using text-to-audio AI models — made for experimental music.

Local. Offline. Open source.

AI Audio Pipeline
01

Interface

Fragmenta App Interface
02

Why?

With AI seemingly everywhere, most tools are either locked behind subscription models designed for consumers, or they require deep technical knowledge, coding skills, and a huge time investment to use effectively. At the same time, the technology's narrative is being driven by big tech, turning access into a commodity. These technofeudal structures limit who gets to shape the future of AI and how it can be used creatively.

There also exist the ethical problems. Much of today's AI is trained on vast amounts of data scraped from the internet, without permission, often infringing on intellectual property rights. Built on open research from Stability AI, Fragmenta shows that ethically trained, personalized models can empower musicians without infringing copyrights or compromising artistic integrity.

Participating in how the narrative of technology is being shaped can be a form of democratic intervention. Fragmenta exists to enable artists to train models on their own work, have a transparent understanding of their AI carbon footprint and use AI on their own terms rather than as a product. Your music or recordings never leave your device.

03

Generated Sounds

Fine-tuned outputs trained on personal audio data.

01

[weird drum beat, 130 bpm]

0:00 / 0:00
02

[arpegio, light, minor, 130 bpm]

0:00 / 0:00
03

[drum beat]

0:00 / 0:00
04

[noisy ambient, texture, full spectrum]

0:00 / 0:00
04

Three Ways to Run

Choose the setup that is accessible to you.

Option A
Hugging Face

Hugging Face Spaces

No installation. Run in your browser. CPU inference — slow, ideal for exploring the interface. Limited GPU sessions available on request.

Open Space ↗
Option B
Docker

Docker

Fastest local setup. No Python or Node.js required. Separate images for NVIDIA GPU and CPU (Mac / Linux / Windows).

docker run -d -p 5001:5001 \
  --gpus all mazcode/fragmenta:gpu
Docker Hub ↗
Option C

Run Locally

Clone the repo and run the launcher. Installs everything in an isolated folder — deleting it removes all dependencies. Linux, Windows, macOS.

./fragmenta.command  # macOS
./fragmenta.sh      # Linux
./fragmenta.bat     # Windows
View on GitHub ↗
05

Four Modules

06

Real-time Monitoring

Training MonitorLive
Epoch
7 / 10
Step
245 / 350
Loss
0.0847
GPU Mem
14.2 GB
Progress70%
Best Loss: 0.0821ETA: 12 minutes
07

Requirements

System
~15GB storage space (including base model)
Internet connection required for installation
Python 3.11 (required for local install — 3.12+ not supported)
CUDA GPU ≥12GB VRAM for LoRA, ≥24GB for full fine-tune
Non-CUDA systems supported for inference (limited performance)
Performance Reference
NVIDIA GPU~3s / 10s audio
Apple Silicon~9m / 10s audio (M1)
HF Spaces CPUvery slow, no setup
Prerequisites
Free Hugging Face account required
Accepting Stable Audio Open T&C
Basic understanding of AI (Recommended)
A considerable amount of your own audio data (Required for fine-tuning)
Note: Fragmenta is intended for experimental music and does not create realistic audio. The project is in active development and not intended for production use. Released under the GNU AGPL v3.0; the underlying Stable Audio models are governed by the Stability AI Community License. Users are solely responsible for ensuring compliance.

This website uses cookies for functionality and analytics (only with your consent). I do not save or sell your personal data. External links may track you through their own services.