Generate stems
Separate a library track into drums, bass, other, and vocals, and keep the cache for later loads.
Offline separation turns any library track that is not already a native NI Stem file into a persistent cache. The cache holds a mixdown plus drums, bass, other, and vocals. Mixar reuses it every time you load the track.
Writing a persistent cache is opt-in. Once the model is on disk, loading a deck can also separate live instead (see Realtime stems on deck load).
Before you start
- The track is in your library and is not a
.stem.mp4. Native Stem files already contain the four layers. - Free disk space is available for one
.stem.mp4per track. - You are not in the middle of a set. Jobs run in the background and do not stop playback, but they compete for CPU, GPU, and disk.
Generation is resource-intensive and can take a while: often on the order of the track’s length, sometimes longer on slower machines. A GPU is recommended when one is available; Mixar will use it automatically. Without a GPU it still runs on CPU, just more slowly.
Generate stems before you perform: prepare tracks in the library ahead of a set so separation does not affect live performance.
Generate one track
- In the library, open the track’s row menu (⋯).
- Choose Generate stems.
- Repeat for any other tracks you want to prepare.
- Wait for the job to finish.
While the job runs
- The menu item becomes Generating stems… and stays disabled until the job finishes.
- The library row can show progress phases (model download, separate, encode).
- If that track is on a deck, the Stems pads show STEMS GENERATING….
- Playback continues.
When it finishes, the Stems pads become available. Generation runs once per track. Mixar reuses the result in later sessions, so the pads are ready as soon as you load the track again. Choosing Generate stems again is safe: Mixar skips the work if a valid cache already exists.
The first generate on a machine may download the separation model into Mixar’s app data. The model chosen in Settings → Library → Stem model (HTDemucs or SCNet int8) is used for new caches. That is a one-time download until you clear the model cache. See Storage for where files land.
Whole collection
- From a collection’s menu, open Analyze.
- Turn on Also generate stems in the dialog.
- Press Start analysis.
Mixar queues stem generation for file tracks in that collection, one at a time. Tracks that already have a valid cache, and native .stem.mp4 files, are skipped.
Realtime stems on deck load
With Settings → Mixer → Realtime Stems enabled, loading a track without a generated cache or native Stem layers starts live StemgenRT separation for that deck. The track begins playing immediately; the Stems pads switch on when the first hop is covered, and the deck crossfades from the original mix. Live stems are never written to disk, so they add no duplicate files. If separation falls behind or you seek outside the prepared window, the deck keeps playing the original mix until it catches up.
Realtime stems run on the CPU and use roughly one core per deck; a GPU is not required. They are a fallback for tracks without offline stems, not a replacement for them. Quality is lower, and Generate stems remains the way to get a persistent, instant-ready cache. The model downloads in the background on first use, and the setting is disabled by default.
When something goes wrong
- Pads disabled, no banner: the track has no cache and Realtime Stems is off or its model has not downloaded yet. Enable it in Settings → Mixer, wait for the download, or generate stems or load a native NI Stem file.
- STEMS GENERATING…: a job is in flight. Wait for it to finish; playback is unaffected.
- Disk full: Mixar stops the write and shows an error toast. Free space, then generate again. BPM/key analysis already done for the track is kept.
To remove a cache and free the disk space, clear it in Settings → Storage: see Storage.