Friendly Stable Audio 3: open-source fork exposes full text-to-audio training pipeline
serrjoa · x · 2026-09-07
Researcher yukara-ikemiya released friendly-stable-audio-3, a refactored training-first fork of Stable Audio 3 for researchers and engineers:
- Full training pipeline: SAME autoencoder, flow-matching diffusion, latent CLAP, and ARC adversarial post-training — not just LoRA fine-tuning
- One generic Trainer: scripts/train.py trains all components
- Config-first: models fully described in YAML, no opaque modelconfig. downloads
- Distributed training via Hydra + HuggingFace Accelerate (multi-GPU/multi-node)
The upstream repo is inference-oriented; this fork exposes the complete training stack.
More from Multimodal
- One-sentence 3D worlds turn cinematic with CineMap and TapNow — gizakdag · 2026-10-26
- No Camera Needed: Creator Builds an Entire World in CineMap and Directs It Into Cinema — gizakdag · 2026-10-26
- Kling vs. Seedance: the Chinese AI video model rivalry — DavidmComfort · 2026-10-26
- Bokeh Cosmic Nebula prompt template: out-of-focus orbs mimic interstellar clouds — LudovicCreator · 2026-10-26
- AIandDesign reveals Runway Act-2 + ElevenLabs voice changer video workflow — AIandDesign · 2026-10-26
- A prompt for generating downhill mountain biking action-sequence AI videos — umesh_ai · 2026-10-26