--- library_name: audio-interv tags: - ace-step - audio - diffusion - interpretability - music - steering - tokemb - token-embedding - vocal-gender --- # Token-Embedding (TokEmb) Direction — `vocal_gender` (ACE-Step) Per-concept T5 hidden-state direction for steering ACE-Step audio generation toward the **vocal_gender** concept. At inference, `TokEmbSteeringController` adds `alpha * direction` to the concept token's hidden state in the neutral prompt embedding before the diffusion process. The `.pt` file is a dict with keys: `direction` (Tensor[hidden_dim]), `concept` (str), `hidden_dim` (int). ## Paper TADA! Tuning Audio Diffusion Models through Activation Steering — [https://huggingface.co/papers/2602.11910](https://huggingface.co/papers/2602.11910) ## Quickstart ```python from src.steering import SteerableACEModel, TokEmbSteeringController model = SteerableACEModel(device="cuda") model.pipeline.load() ctrl = TokEmbSteeringController.from_pretrained( "lukasz-staniszewski/ace-step-tokemb-vocal-gender", alpha=1.0, te_split_step=3, ) with model.steer(ctrl): audio = model.generate( prompt="instrumental music", lyrics="[inst]", audio_duration=10.0, infer_step=30, manual_seed=0, ) ``` ## Metadata ```json { "hidden_dim": 768, "direction_file": "vocal_gender_direction.pt" } ```
