LoRA in Stable Audio 3

LoRA fine-tuning lets you adapt a Stable Audio 3 model to a specific style, sound, or domain without retraining the whole model. The result is a small `.safetensors` file (~50–200 MB) that you load on top of any base checkpoint at inference time — stackable, adjustable in strength, and swappable without touching the base weights. - A dataset of audio files with matching text descriptions (at minimum ~20–50 clips; more is better) - A CUDA GPU with sufficient VRAM: | Model | Standard | With `--base_precision bf16 --adapter_type lora-xs` | |---|---|---| | `medium` | ~6.5 GB | ~5.5 GB | | `small` | ~2.5 GB | ~2 GB | - The `lora` extra installed: `uv sync --extra lora` We don't claim these are optimal settings, LoRA behavior varies a lot with dataset size, style, and hardware. But these are the configurations we've found work well for most datasets and are good starting points before tuning. Good default for most datasets. `dora-rows` is the default adapter and tends to generalize well.