Dayglass Separate
=================

Splits a track into stems on your own machine. Nothing is uploaded, and nothing
is charged.

This is the same code the Dayglass Audio GPU runs: same model, same refinement
pass. The stems come out the same. The difference is time, because a processor
is doing work a graphics card is built for.


BEFORE THE FIRST RUN
--------------------

Install uv, a Python runner. Open a terminal and run:

    winget install astral-sh.uv

Then close that terminal. Nothing else needs installing: uv fetches Python,
PyTorch, and the separation model itself.


SEPARATING A TRACK
------------------

Drag an audio file onto "Dayglass Separate.cmd".

Choose how careful to be when it asks. Balanced is the sensible default. Ultra
does four passes instead of two and takes about twice as long.

Then wait. The first run downloads a few hundred megabytes and takes a while.
After that, a four-minute track takes roughly five to fifteen minutes depending
on your processor.

The stems land in a folder next to the track, named after it.


USING THEM
----------

Open dayglassaudio.com, load the same track, and choose MY OWN STEMS. Point it
at the folder this wrote.

From there it works exactly as though the separation had happened on the site:
the same measurements, the same patch, the same export. No upload, no challenge,
no daily limit, no token.


WHAT YOU GET
------------

Up to six files: drums, bass, vocals, guitar, piano, other.

Instruments the track does not contain are left out rather than written as
near-silence, so a song with no piano gives you no piano file rather than one
holding separation residue.


IF IT WILL NOT START
--------------------

"This needs uv"
    Install it with the command above, then open a NEW terminal window. An
    already-open one will not see it.

It runs but takes forever
    That is normal on a processor. Try fast mode first to watch it finish.

Anything else
    support@dayglassaudio.com
