metadata
library_name: audio-interv
tags:
- activation-steering
- audio
- austeer
- diffusion
- interpretability
- music
- steering
- tf6tf7
- violin
AUSteer — violin (ACE-Step)
Per-(step, layer) sparse activation-momentum scores for the violin concept on ACE-Step. At inference, AUSteerSteeringController adds alpha along the top-k most concept-discriminative bins.
Paper
TADA! Tuning Audio Diffusion Models through Activation Steering — https://huggingface.co/papers/2602.11910
Quickstart
from src.steering import SteerableACEModel, AUSteerSteeringController
model = SteerableACEModel(device="cuda")
model.pipeline.load()
ctrl = AUSteerSteeringController.from_pretrained(
"lukasz-staniszewski/ace-step-austeer-violin-tf6tf7", alpha=15.0, k=256, mode="additive",
)
with model.steer(ctrl):
audio = model.generate(
prompt="instrumental music", lyrics="[inst]",
audio_duration=10.0, infer_step=30, manual_seed=0,
)
Generation config
{
"method": "austeer",
"concept": "violin",
"lyrics": "[inst]",
"layers": "tf6tf7",
"layers_collected": [
"tf6",
"tf7"
],
"num_inference_steps": 30,
"audio_duration": 30.0,
"seed": 10,
"guidance_scale": 5.0,
"guidance_scale_text": 0.0,
"guidance_scale_lyric": 0.0,
"guidance_interval": 1.0,
"guidance_interval_decay": 0.0
}