Eisbach-Medium

A belief-space fine-tuned LoRA adapter for Stable Audio 3 Medium.

Eisbach-Medium is a DoRA-rows adapter (rank 16, 41MB) fine-tuned with the Eisbach log-barrier — a self-referential confidence-gating mechanism that reshapes training dynamics to favor temporally structured, acoustically differentiated outputs. It specializes in long-form instrumental music with clear structural development and narrative coherence.

Quick start

This is a LoRA adapter — you need the base Stable Audio 3 Medium model.

Recommended parameters

Parameter	Recommendation
Steps	8 (fast) to 50 (refined)
CFG scale	2.5–4.0
Duration	30s–180s
Sampler	pingpong
Prompt style	Descriptive, literary, scene-based

Training

Field	Value
Base model	Stable Audio 3 Medium (2.3B DiT, 44.1kHz stereo)
Adapter	DoRA-rows, rank 16, alpha 16
Steps	1,000
Batch size	4
LR	5e-5 (AdamW)
Dataset	MusicCaps
Eisbach barrier	0.5

Limitation

• Instrumental music only; vocals not tested
• Favors chamber/orchestral textures over loop-based genres
• Scene-based prompts work better than technical tag lists

License

Derived from Stable Audio 3 Medium. Subject to Stability AI's model license terms.