sa3-drone-bender / README.md
Danhearn's picture
Point to maxgraf/sa3-variations-models for base weights; this repo hosts only bending directions
0fdb16a verified
|
Raw
History Blame Contribute Delete
2.17 kB
metadata
license: other
license_name: stability-ai-community
license_link: https://stability.ai/license
tags:
  - mlx
  - audio
  - music-generation
  - stable-audio
  - text-to-audio

SA3 Drone Bender — bending direction vectors

Contrastive bending direction vectors for the SA3 Drone Bender JUCE plugin: a real-time contrastive-DiT-bending + drift/feedback-loop instrument built on a C++/MLX port of Stable Audio 3.

Files

File Size Description
bending_directions_medium.safetensors 244K 5 contrastive bending direction vectors (layers 4-11 of the medium DiT)

5 semantic direction vectors (Harmonic Brightness, Melodic Activity, Textural Density, Rhythmic Intensity, Register) extracted via contrastive analysis of the medium DiT's internal activations (layers 4-11), used to steer generation in the Drone Bender's "Bend" XY pad.

Base model weights

This repo does not host the base Stable Audio 3 weights. Get those from maxgraf/sa3-variations-models — an MLX/safetensors port of Stability AI's stabilityai/stable-audio-3-optimized medium bundle: dit_medium_f16.safetensors, same_l_encoder_f32.safetensors, same_l_decoder_f32.safetensors, t5gemma_f16.safetensors.

bending_directions_medium.safetensors must sit alongside dit_medium_f16.safetensors in the same models directory — the orchestrator loads it as a sibling file.

Usage

The setup.sh script in the plugin repo downloads all five required files (four from maxgraf/sa3-variations-models, this one from here) and builds the plugin in one step. To fetch just this file manually:

hf download Danhearn/sa3-drone-bender bending_directions_medium.safetensors \
  --local-dir /path/to/models

License

bending_directions_medium.safetensors is original work derived from Stable Audio 3's medium DiT internal activations, released under the Stability AI Community License for consistency with the base model weights.