metadata
library_name: audio-interv
tags:
- ace-step
- activation-steering
- audio
- austeer
- diffusion
- interpretability
- mood
- music
- steering
Steering Vectors — mood (ACE-Step)
Steering vectors for the mood concept on ACE-Step, computed via contrastive activation addition (CAA).
Quickstart
from src.steering import SteerableACEModel
from steering.caa import CAAController
model = SteerableACEModel(device="cuda")
ctrl = CAAController.from_pretrained("lukasz-staniszewski/ace-step-austeer-mood-tf6tf7", alpha=20)
with model.steer(ctrl):
audio = model.generate(prompt="instrumental music", seed=0)
Generation config
{
"method": "austeer",
"concept": "mood",
"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
}