Status: parked. Working code, wrong time. It needs a locally running ACE-Step 1.5 server (CUDA GPU) to do anything, and the AI music channel it was built for is a back-burner project. Kept here as a record of the engineering — FastAPI backend, real audio pipeline, honest scope.
A local-first AI music workstation built around your two core ideas:
- Kinetic Flow: still → moving, rather than quiet → loud.
- Lock-In Loop: repeatable, predictable cycles where anticipation and payoff happen in a familiar place.
FlowForge does not require a paid music API. It connects to a locally running ACE-Step 1.5 server.
- Local text-to-music generation through ACE-Step 1.5
- Kinetic Flow, Lock-In Loop, Cinematic Tension and Custom presets
- Controls for movement, tension, rhythmic density and predictability
- Instrumental-first prompting
- 1–4 minute master generation
- WAV / MP3 / FLAC output selection
- Optional BPM and seed control
- In-app job polling and playback
- 10, 30 and 60 minute repeated loop export via ffmpeg
ACE-Step's current official project exposes a local REST API. The default FlowForge connection is:
http://127.0.0.1:8001
Typical ACE-Step setup:
git clone https://github.com/ACE-Step/ACE-Step-1.5.git
cd ACE-Step-1.5
uv sync
uv run acestep-api
ACE-Step will download model files on first use. A CUDA GPU is recommended, but ACE-Step also documents other backends.
If you run ACE-Step on another URL, set ACE_STEP_URL before launching FlowForge.
Windows example:
set ACE_STEP_URL=http://127.0.0.1:8001
run_windows.batDouble-click:
run_windows.bat
./run_mac_linux.shThen open:
http://127.0.0.1:8090
Generation works without ffmpeg. Long-loop export uses ffmpeg.
On Windows, install ffmpeg and make sure ffmpeg.exe is on PATH.
For cinematic, choreographic instrumentals: tension becomes motion. The payoff is fluidity and forward movement, not just loudness.
For repeatable focus/training/background tracks. The cycle is intentionally predictable and the ending is prompted to reconnect naturally with the beginning.
For patient dramatic underscore with negative space and an earned release.
Do not feed copyrighted soundtrack audio into reference/cover modes unless you have the rights to do so. FlowForge's default workflow uses abstract musical qualities and original prompts rather than copying a source track.
- True seamless-loop editor with waveform trim + crossfade
- Reference-analysis mode for audio you own
- "Payoff position" and "cycle length" controls
- Track history/library with favorite generations and seeds
- Side-by-side A/B variations from the same seed
- Stem separation and per-stem regeneration
- LoRA/style-profile support for your own legally controlled training material
- Desktop packaging so it launches like a normal app