AnchorSteer Pretrained Weights
Pretrained Controller checkpoints for the KDD 2026 paper:
AnchorSteer: Self-Discovered Concept Injection for Structure-Preserving Music Editing
- GitHub: hengtsune1024/AnchorSteer
- Code archive: 10.5281/zenodo.20368360
Contents
Each checkpoint directory contains:
| File | Description |
|---|---|
best.pth |
Controller weights at the epoch with lowest validation loss |
config.yaml |
Full training configuration for reproducibility |
This release includes 5 concepts:
| Directory | Concept type | Concept |
|---|---|---|
exps/ambient-transformer/ |
Genre | Ambient |
exps/harp-transformer/ |
Instrument | Harp |
exps/jazz-transformer/ |
Genre | Jazz |
exps/piano-transformer/ |
Instrument | Piano |
exps/rock-transformer/ |
Genre | Rock |
All checkpoints use the Transformer Controller architecture trained on full 47-second audio (20 epochs, lr=5e-4, DiT blocks 0–23).
Usage
1. Install AnchorSteer
Follow the installation instructions in the GitHub repo.
2. Download checkpoints
# from the AnchorSteer root directory
hf download heng1024/AnchorSteer-weights --repo-type model --include "exps/*" --local-dir .
This places the exps/ folder directly into the repo root, matching the expected layout.
3. Run inference
# Generate concept-steered audio from a text prompt
python generate.py --exp_dir exps/rock-transformer
# Structure-preserving editing of a source file
python edit.py --source_audio path/to/source.wav --concept_dir exps/rock-transformer
See the README for full options.
Citation
@inproceedings{anchosteer2026,
title = {AnchorSteer: Self-Discovered Concept Injection for Structure-Preserving Music Editing},
author = {Chih-Heng Chang, Keng-Seng Ho, Chih-Yu Tsai, Kuan-Lin Chen, Yi-Hsuan Yang, Jian-Jiun Ding},
booktitle = {Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2},
year = {2026},
}
License
MIT — see LICENSE.