Localsong: 1.2B Parameter Game Music Model Trained on One H100
Amazing-You9339 · reddit · 2026-08-23
The author released Localsong, an instrumental game music generator model.
Model & Training Details:
- Architecture: 1.2B parameter DiT model.
- Resources: Trained from scratch on a single cloud H100 GPU over 8 days.
- Components: Uses the VAE from Stable Audio 3.
- Goal: Covers a wider range of instrumental styles than Ace-Step, Minimax M3, or Stable Audio 3 (no lyrics).
- Open Source: The repo includes a WebUI and MP3 samples, runnable via uv run webui.py.
Related event: 1.2B Game Music Generation Model Trained on a Single H100 Goes Open Source(2 posts)→
More from Multimodal
- Flashback: exploring immersive virtual worlds in latent space via prompts — nptacek · 2026-08-23
- H3 Video Generation Test: Why Does Character Control in Complex Scenes Rely on Luck? — Hdfjds · 2026-08-23
- Minimax Leads in Prompt Adherence, Flux 3 Wins on Atmosphere and Texture — Grinderius · 2026-08-23
- One Person, Three Weeks: AI-Generated Sci-Fi Film SPECTRUM Released — HashemGhaili · 2026-08-23
- Short film created using Minimax H3 audio and ComfyUI r2v workflow — Repulsive-Rush3505 · 2026-08-23
- Sora-level VFX Spellcasting: Emissive Runes and Zero-G Physics — azed_ai · 2026-08-23