TADA Steering Collection
Collection
Steering vectors, SAEs, and counterfactual prompts for ACE-Step. Companion to the audio-interv unified Controller interface. • 87 items • Updated • 2
SAE — ACE-Step (transformer_blocks.7.cross_attn)
Sparse autoencoder trained on ACE-Step cross-attention activations.
Paper
TADA! Tuning Audio Diffusion Models through Activation Steering — https://huggingface.co/papers/2602.11910
Quickstart
from src.steering.methods.sae.lib.sae.sae import Sae
sae = Sae.load_from_hub(
"lukasz-staniszewski/ace-step-sae-tf7-cross-attn",
hookpoint="transformer_blocks.7.cross_attn", device="cuda",
)
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support