Open Source Stereo-to-Spatial Audio Model
kittenkrazy · reddit · 2026-07-18
The author open-sourced Stereo2Spatial, a model that converts stereo music into spatialized binaural mixes, releasing training/inference code, a Hugging Face model, and a Windows desktop app.
Key Implementation
- An initial VAE latent-based approach hit an audio quality bottleneck due to mismatched output distributions
- The author switched to directly modeling raw waveforms, which is more compute-intensive but yields better results
- Training instability was ultimately solved via amplitude lifting: normalizing tracks to an RMS of 0.33 before multiplying them back to expand the model's space
Training & Release
- Trained on 7,669 tracks over 20 days using 2× A6000 GPUs
- Two-stage training: short sequences for the first 10 days, long sequences for the latter 10 days
- Supports optional mix-style conditional control
- The binaural output version is released first, with a 7.1.4 version planned for the future pending more compute
- Open-sourced under Apache 2.0
Links
- Hugging Face: stereo2spatial-v2-binaural and stereo2spatial-v1
- GitHub: Training/inference repos and the Windows App are public
- Includes a detailed case study documenting the shift from the latent to the waveform approach
More from Multimodal
- Creator builds Snorricam video series with Seedance 2.5, sharing prompts — LudovicCreator · 2026-10-27
- Gradium launches free webpage TTS Chrome extension powered by its 40M-parameter Phonon model — mattturck · 2026-10-27
- Vortex of Infinite Recursion prompt goes viral: images that spiral endlessly into themselves — azed_ai · 2026-10-27
- 170+ AI Creators Made One Collective Film, Premiering November 6 — Loo_Atreides · 2026-10-27
- ComfyUI on 12GB VRAM: measured peaks show offloading turns 24s into 11 minutes — Impossible-Mail-4619 · 2026-10-27
- Midjourney creator showcases --sref style reference outputs: gothic character art — OVolosin82152 · 2026-10-27