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

python

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

json

{
  "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
}

lukasz-staniszewski/ace-step-austeer-violin-tf6tf7

AUSteer — violin (ACE-Step)

Paper

Quickstart

Generation config

lukasz-staniszewski/
ace-step-austeer-violin-tf6tf7

AUSteer — `violin` (ACE-Step)