Tutorial · Python · ~10 min
Demucs Alternative (Cloud, No GPU Needed)
You know Demucs. You also know the setup: PyTorch, CUDA, VRAM limits, and a job that dies at midnight. Here is the cloud version.
Why this matters
Demucs/HTDemucs produces great stems but demands GPU infrastructure and babysitting. For volume work — catalogs, batches, pipelines — the setup cost never amortizes.
A managed API runs the same class of models on shared GPU workers: upload a URL, poll, get stems. No environment, no VRAM math.
How it works
Submit the track with your stem list; the API returns each stem as a separate WAV/FLAC, time-aligned and ready to mix.
Submit a job with an input URL (S3, GCS or HTTPS), poll the job URL, and download the rendered output. No GPU, no queues, no ffmpeg builds to babysit.
Code
import requests, time
API = "https://api.mlslabs.io/v1/source-separation/jobs"
headers = {"X-API-Key": "YOUR_API_KEY"}
payload = {"input": "s3://bucket/song.wav", "stems": ["drums", "bass", "vocals", "other"]}
resp = requests.post(API, json=payload, headers=headers)
job = resp.json()
while job["status"] not in ("succeeded", "failed"):
time.sleep(3)
job = requests.get(job["url"], headers=headers).json()
print("Output:", job["output_url"])Pro tips
- Compare the API's output to your local Demucs on the same 3 songs before migrating.
- Lossless input (WAV/FLAC) separates better than MP3.
- Batch the catalog after the A/B — volume is where the API wins.
Pricing note
Usage is metered per minute of media processed; the first tier is free each month. Volume discounts kick in automatically.
FAQ
Common questions
Is the model the same as Demucs?
We use BS-RoFormer-class models (a newer family); the honest answer is to A/B on your own tracks.
Do I keep stems in sync?
Stems are sample-aligned by default — drop them straight into a DAW.
What about very long tracks?
Hours-long files are chunked and stitched; stems remain perfectly aligned.