Skip to content

Generating with a LoRA

This example generates new music from a LoRA trained on one album, using Field Audio. The checkpoint, a rank-16 Underfit LoRA, trained for 1,000 steps on 165 lossless 12-second chunks from all 12 tracks on Henry Mancini's Experiment in Terror, is applied to Stable Audio 3's small-music base model.

To run the examples on this page, open the audio_ai_underfit_mancini starting point.

The source is a 12-second excerpt of the Modern Jazz Quartet’s “Skating”, separate from the Mancini album used for training. The request imports that excerpt and asks for a 30-second candidate with a modest guidance value (cfg = 1.5). initNoiseLevel = 0.65 sets how far the result may move from the source; the LoRA applies at strength 0.85 over the denoising interval 0..0.92. The candidate board records the source, seed, model revisions, checkpoint and every sampling control, so a take can be reproduced.

Listen to the experiment

Listen to the source first, then compare the base model with the LoRA take. Both outputs use seed 1962, 50 steps, cfg = 1.5, cfgRescale = 0, initNoiseLevel = 0.65, two seconds of padding and chunked decoding. The only request parameter that changes is checkpoint strength: zero for the baseline and 0.85 for the LoRA, so the difference you hear is the checkpoint's.

Source · MJQ, “Skating” · 12 seconds

Stable Audio 3 Small Music Base · no LoRA · 30 seconds

Mancini step-1000 checkpoint · strength 0.85 · 30 seconds

Actual MJQ source waveform and Underfit candidate boards for the matched no-LoRA baseline and original Mancini LoRA take

The board above shows the source waveform and both takes. In a live document you audition and rate candidates on the board itself.

Where these examples came from

The LoRA take is the successful 13 September 2026 render retained on the Linux RTX 4090 machine, not a new replacement. Its native WAV hash is 572545c7809e33ad00e642d40c5b3c65ebecfa3abbf233b8924292e26d30a5ad. The matched strength-zero baseline was rendered on 3 October 2026 with the same pinned Underfit and Stable Audio 3 revisions and request controls. The checkpoint digest is 3769f0e3282ea970a5c953933216f84c847478cfc51213c7adfa3f5a98762364. Candidate proof, baseline proof and the export manifest preserve the original asset hashes, model revisions and encoding provenance. Opus copies use the real audition proxies without gain normalization. The captured board omits the unavailable archived splice preview; the source is provided separately. Refresh these exports before recapture; a normal HTML build reuses the genuine results and does not run a model.

Run it

Start the isolated worker and API from the field-audio-ai checkout:

export FIELD_AUDIO_UNDERFIT_SOURCE_WAV=/path/to/your-source.wav  # at least 12 seconds
export FIELD_AUDIO_UNDERFIT_LORA_ROOT=/path/to/audio-project/underfit-loras
./scripts/run-underfit-worker.sh
./scripts/run-api.sh

The worker also needs the Stable Audio 3 base model, found through FIELD_AUDIO_UNDERFIT_MODELS_ROOT. Then open the Underfit Mancini example: its first box explains the setup, and its second imports the source, sends the LoRA request and shows the candidates on a board you can listen through.

See also