Developer Releases Open-Source 1.2B Diffusion Transformer for Game Music Generation
A developer has trained a 1.2B parameter Diffusion Transformer (DiT) model named "Localsong" from scratch to generate instrumental game music. The model was trained on a single cloud H100 GPU over eight days, and the weights, WebUI, and audio samples have been released on Hugging Face. This release provides game developers and creators with a local, open-source alternative for generating high-quality instrumental music across diverse styles without relying on proprietary APIs. It demonstrates that training capable, specialized audio generation models is increasingly feasible for individual developers using consumer-accessible cloud hardware. The model utilizes the Variational Autoencoder (VAE) from Stable Audio 3 and focuses strictly on instrumental tracks without lyrics. Users can run the provided WebUI locally using the fast Python package manager `uv` with the command `uv run webui.py`.
## BACKGROUND
Diffusion Transformers (DiTs) combine the generative capabilities of diffusion models with the scalable transformer architecture, which is widely used in state-of-the-art AI models. Stable Audio uses a Latent Diffusion Model (LDM) architecture where a Variational Autoencoder (VAE) compresses audio into a compact latent space for efficient processing. The `uv` tool is a modern, extremely fast Python package manager written in Rust that simplifies dependency management and project execution.