StemFX: Learning Mixing Style Representations via Autoregressive FX Chain Prediction on Source-Separated Stems

Yuan-Chiao Cheng, Jui-Te Wu, Brian Chen, Yen-Tung Yeh, Yu-Hua Chen, Yi-Hsuan Yang

The official implementation of the ISMIR 2026 paper "StemFX: Learning Mixing Style Representations via Autoregressive FX Chain Prediction on Source-Separated Stems".

Paper  |  Demo  |  Weights

StemFX is a paired encoder–decoder for music mixing:

• BSFiLM Encoder — Band-split CNN with FiLM conditioning + temporal attention pooling. Maps a 4-stem 10-second audio segment to a single L2-normalized 512-d embedding capturing mixing style.
• FX Chain Generator — A 6-layer transformer decoder that, given a pair of (original, target) embeddings, autoregressively emits a per-stem FX chain in multiafx format. The chain can be applied to the original stems to render a new mix in the target style.

The two are trained jointly on a Sep-Aug Pipeline: source-separated FMA stems are randomly remixed across songs and processed through random multilib FX chains; the model learns to recover the chain that explains the observed difference.

pip install stemfx

For training and evaluation:

pip install "stemfx[training,evaluation]"

Separating arbitrary mixtures needs no extra setup. The SCNet source is bundled, and its pretrained weights (~214 MB) are downloaded from the upstream project's release on first use and cached under ~/.cache/stemfx (override with STEMFX_CACHE_DIR).

import stemfx
import soundfile as sf
import torch

model = stemfx.load()                          # downloads the paper checkpoint from HF Hub

# 1a) Embedding a mixture — separated internally with SCNet.
#     The separator checkpoint is fetched and cached on first use.
emb_a = model.embed("song_a.wav")              # → (512,) L2-normalized
emb_b = model.embed("song_b.wav")

# 1b) Or pass pre-separated stems directly (skips the separator entirely).
stems = {
    name: torch.from_numpy(sf.read(f"{name}.wav", dtype="float32", always_2d=True)[0].T)
    for name in ("vocals", "bass", "drums", "other")
}
emb = model.embed(stems)

# To use your own separator checkpoint instead of the default:
model = stemfx.load(scnet_model="path/to/scnet.ckpt",
                    scnet_config="path/to/scnet.yaml")

# 2) Predicting an FX chain from a pair of embeddings (multiafx-format)
chain = model.transfer(emb_a, emb_b)
# chain == {"vocals": [{"effect": "...", "params": {...}}, ...],
#          "bass": [...], "drums": [...], "other": [...]}

# 3) End-to-end audio: render the chain on song_a's stems
result = model.transfer_audio("song_a.wav", "song_b.wav")
print(result.pretty())                         # human-readable chain summary
torchaudio.save("transferred.wav", result.mix, result.sample_rate)

TransferResult exposes:

python pipeline/preprocess_fma.py \
    --input_dir /path/to/fma_full/ \
    --output_dir /path/to/fma_separated/ \
    --scnet_model third_party/Music-Source-Separation-Training/model_scnet_masked_ep_111_sdr_9.8286.ckpt \
    --scnet_config third_party/Music-Source-Separation-Training/configs/config_musdb18_scnet_xl_ihf.yaml \
    --device cuda:0

python pipeline/augment_stems.py \
    --source /path/to/fma_separated/ \
    --output /path/to/fma_separated_augmented/ \
    --min_fx 1 --max_fx 10 \
    --workers 32

Required for cross-song stem mixing during training (mix_stems: true):

python pipeline/compute_stem_lufs.py \
    --dataset_dir /path/to/fma_separated_augmented/ \
    --threshold -50.0 \
    --num_workers 32

python train.py \
    --config configs/default.yaml \
    data.original_path=/path/to/fma_separated \
    data.augmented_path=/path/to/fma_separated_augmented \
    train.checkpoint_dir=checkpoints/stemfx_main \
    train.log_dir=logs/stemfx_main \
    train.device=cuda:0

CLI dotted overrides take precedence over the YAML file. See configs/default.yaml for the full set of knobs.

python evaluate/retrieval.py \
    --dataset /path/to/MUSDB18_paired_fx_corrected_5/ \
    --checkpoint checkpoints/stemfx_main/best_checkpoint.pt \
    --device cuda:0

python evaluate/style_transfer.py \
    --original_dir /path/to/MUSDB18_normalized/test \
    --target_dir /path/to/MUSDB18_styled/test \
    --checkpoint checkpoints/stemfx_main/best_checkpoint.pt \
    --output_dir outputs/eval_st \
    --num_examples 50

stemfx/                       # installed Python package
    api.py                    # high-level inference wrapper
    encoder.py                # BSFiLM Encoder
    generator.py              # FX Chain Generator
    model.py                  # StemFX nn.Module (encoder + generator)
    tokenizer.py              # FX-chain tokenization (multiafx vocabulary)
    losses.py                 # FXChainLoss, MultiResolutionSTFTLoss
    separator.py              # SCNetSeparator (bundled SCNet)
    _scnet_infer.py           # chunked overlap-add separation
    weights.py                # checkpoint resolution + download
    third_party/              # vendored SCNet source (MIT, see NOTICE)
    mixing_features.py        # FiLM feature extractor (44.1 kHz)
    normalization.py          # multiafx parameter ranges

configs/default.yaml          # training config (paper main)
data/sepaug_dataset.py        # Sep-Aug paired dataset + DataLoader
pipeline/                     # data preprocessing scripts
    preprocess_fma.py         # SCNet separation
    augment_stems.py          # multilib FX augmentation
    compute_stem_lufs.py      # per-stem LUFS + valid_tracks.json
evaluate/                     # evaluation scripts
    retrieval.py              # paired FX retrieval
    style_transfer.py         # free + forced MRSTFT
train.py                      # config-driven training entry point
tests/                        # pytest smoke tests

@inproceedings{cheng2026stemfx,
  title     = {{StemFX}: Learning Mixing Style Representations via Autoregressive
               {FX} Chain Prediction on Source-Separated Stems},
  author    = {Cheng, Yuan-Chiao and Wu, Jui-Te and Chen, Brian and
               Yeh, Yen-Tung and Chen, Yu-Hua and Yang, Yi-Hsuan},
  booktitle = {Proceedings of the 27th International Society for Music
               Information Retrieval Conference (ISMIR)},
  address   = {Abu Dhabi, UAE},
  year      = {2026},
  eprint    = {2607.15634},
  archivePrefix = {arXiv},
  primaryClass  = {eess.AS},
}

• Code (this repository): MIT.
• Paper and accompanying documentation: CC BY 4.0.