FLOMSS: flow matching splits music into 4 stems and takes 2nd place in listening tests

For a decade, music source separation has been the territory of discriminative models: a network looks at the mix and directly predicts which part is vocals and which is bass. Researchers led by Shuyang Cui have proposed FLOMSS, a system that treats the same task as generation. The model builds stems with flow matching for full-band stereo in a four-source setup: vocals, bass, drums and "other." The larger FLOMSS-M took second place overall in the listening test, but it trailed strong discriminative separators on SDR.

Four design choices, tested one by one

Rather than showcase a single successful architecture, the authors broke the design space into its components. They compared what the model predicts, which path the flow takes from noise to data, how best to handle the sources, and how to make the stems sum back to the original mix.

• prediction targets;
• flow paths;
• modeling a single source versus modeling all four jointly;
• mixture consistency constraints.

The winning combination was Optimal Transport-CFM with x-prediction. It gave the most balanced result across the full set of metrics. It was not the best on every number but the least lopsided, which is a practical property for a generative model judged by several quality criteria.

SDR is no longer the only yardstick

SDR, the signal-to-distortion ratio, measures how closely the output matches the reference at the waveform level. That suits discriminative models, which optimize for exactly this kind of match. A generative model can produce a convincing drum part that sounds right but does not match the original in phase, and SDR will penalize it.

For that reason, the authors evaluated separation with five groups of metrics:

• signal-level metrics (including SDR);
• spectral metrics;
• embedding-based metrics;
• distributional metrics;
• subjective listening tests.

The result revealed a trade-off between distribution matching and reconstruction. Flow matching models improve spectral and embedding similarity but tend to give up some SDR. For listeners, and for tasks like remixing or sampling, that can be a good deal. For tasks that demand bit-level accuracy, it is not.

Two models and an honest second place

On the MUSDB18-HQ benchmark, the authors compared two configurations.

• FLOMSS-S outperforms earlier generative baselines on most spectral and perceptual metrics.
• FLOMSS-M, the scaled-up version, improves on FLOMSS-S across all reported objective metrics and takes second place overall in the listening test. Its SDR is lower than that of strong discriminative separators.

This is a cautious win. The generative approach has not displaced classical methods, but it comes close to them in perceived quality. "Second place" says more than a record on a single metric: listeners do not find the result fundamentally worse than the leader's, even though its reconstruction numbers are weaker.

Who needs a generative separator

For developers of audio tools, the main takeaway is to choose metrics to fit the task. If a product is meant for karaoke, remixing or preparing datasets for generative music models, it makes sense to look at spectral and embedding metrics and at listening tests, not only at SDR. If you need to subtract vocals from a master track without losing detail, discriminative models are still more reliable.

For researchers, the work offers a ready-made map of design decisions: OT-CFM with x-prediction as a starting point, and a set of metrics that shows the side of quality where SDR is blind. The authors have published a demo page with separation examples. The open question is whether further scaling can close the SDR gap, since the step from S to M already improved all objective metrics.