ACE-Step Audio Steering Suite
Collection
Steering vectors, SAEs, and counterfactual prompts for ACE-Step. Companion to the audio-interv unified Controller interface. • 83 items • Updated • 2
ACE-Step Concept Slider — vocal style (r=16, tf6tf7)
Concept Sliders LoRA for steering vocal style in ACE-Step, trained at LoRA rank 16 on the functional layers tf6-tf7 (lr=1e-4, eta=7, 500 iterations, 100 prompts). Rank selected per concept on a held-out prompt set (see the benchmark paper). Part of the ACE-Step Audio Steering Suite.
from src.steering import ConceptSlidersSteeringController
ctrl = ConceptSlidersSteeringController.from_pretrained("lukasz-staniszewski/ace-step-cs-vocal-style-r16-tf6tf7", alpha=0.1)
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support