The single ZeroGPU API behind Kiku AI, built by Reubencf. The frontend is a separate CPU Docker Space; ZeroGPU only supports the Gradio SDK, so all inference lives here.
| Endpoint | Does |
|---|---|
/accompany | Writes a band under your real recording, following its beat and chords, and keeps your voice |
/part | One instrument (drums, bass, keys, guitar, strings, synth) that follows a mix, or from a prompt |
/transcribe | Audio to melody, chords, ABC, MIDI and a printable score |
/song | A full song with generated vocals from style and lyrics |
/plan_score | YuE2's symbolic plan without audio; returns editable native ABC and a seed |
/score_instrumental | Render a reviewed score as an instrumental, transferring the sung melody into Ins |
/cover | Re-arranges a transcribed melody in a new style |
/instrumental | A backing track from a description |
/remix /repaint /extend /edit | Reshape existing audio; /repaint takes fractional seconds |
/stems | Split a mix into vocals and instrumental |
/split | Split a mix into vocals, drums, bass and everything else |
/score | Reads a stretch into a score: bars with their start times, chords, sections, key, tempo and (optionally) the sung words |
/lyrics | The sung words as plain lines, written down by HeartTranscriptor from the separated vocal |
/restage | Covers a stretch with YuE2: the same melody and chords, sung or on an instrument, locked to the original bars |
/band_score | A YuE2 band that plays the take's own chords and form, locked to its bars, under the real voice |
/score, /restage and /band_score work through the score rather than the sound.
SheetSage2 reads the melody, chords, sections and every beat of a stretch; scores.py edits
that score (chords per bar, transposition, tempo, giving the vocal line to an instrument,
removing it for a band) and rewrites it in the native two-voice ABC, checked with the official
abc_tools.py parser (vendored, MIT). YuE2 then plays the score, so the tune is the original's.
YuE2 has no timing or audio-reference input, so sync.py reads its render again, pairs its bars
with the score's bars by harmony (dynamic time warping, tolerant of a bar added or dropped), and
Rubber Band stretches it until every bar starts where the recording's bar starts. A new tempo
disables that lock, since the original bar lines no longer apply.
No model re-sings you. /accompany writes a band and mixes it underneath your audio.
A tempo and key in the prompt are only a hint to ACE-Step, so a band written from text alone starts on its own downbeat, drifts and plays its own chords. It follows audio closely, though. So the engine:
separate="auto"). If there is some, the voice
is lifted off, and the old backing is what gets analysed: it carries the beat and harmony far
more reliably than a melody does.strength sets how closely the band keeps the guide's timing and chords.split_parts it also
returns drums, bass and the rest as separate stems.It returns the mix, the band, your vocal, the prompt, the original backing (when one was found), the three band stems and a JSON summary of the tempo, key, beats and chords it used.
Singing2Accompaniment is still unreleased, so the band follows your beat and chords, not your
phrasing. Rubato takes keep their tracked beats instead of a fitted grid, but unmetered singing
with no clear pulse falls back to text-to-music at the measured key.
python test_band_tools.py checks the analysis on synthetic audio with a known tempo and
progression; python test_audio_tools.py checks the mixing.
They pin conflicting versions — transformers 4.45.2 (SheetSage2), 4.50.0 (ACE-Step) and 4.57.6
(YuE2) — so pip cannot satisfy all three. The YuE2 wheel's stack is installed and the other two
run on it; ACE-Step is installed with --no-deps and its inference dependencies are listed
explicitly in requirements.txt.
Each model is built on first use, not at import. YuE2 requires this: it reaches low-level CUDA
init while constructing, which ZeroGPU refuses outside a @spaces.GPU function. Weights are
fetched at startup, so no request pays to download them, but the first call to each model
carries its load time and its GPU budget allows for that.
Requests are queued one at a time, uploads are capped at 100 MB, and temporary results expire after 24 hours. Call with the signed-in user's Hugging Face token to use that user's ZeroGPU allowance.
SheetSage2 and YuE2 are CC BY-NC 4.0 (non-commercial). ACE-Step is Apache 2.0 and Demucs is MIT,
so /accompany, /part, /instrumental, /stems and /split results carry no non-commercial restriction.
The single ZeroGPU API behind Kiku AI, built by Reubencf. The frontend is a separate CPU Docker Space; ZeroGPU only supports the Gradio SDK, so all inference lives here.
| Endpoint | Does |
|---|---|
/accompany | Writes a band under your real recording, following its beat and chords, and keeps your voice |
/part | One instrument (drums, bass, keys, guitar, strings, synth) that follows a mix, or from a prompt |
/transcribe | Audio to melody, chords, ABC, MIDI and a printable score |
/song | A full song with generated vocals from style and lyrics |
/plan_score | YuE2's symbolic plan without audio; returns editable native ABC and a seed |
/score_instrumental | Render a reviewed score as an instrumental, transferring the sung melody into Ins |
/cover | Re-arranges a transcribed melody in a new style |
/instrumental | A backing track from a description |
/remix /repaint /extend /edit | Reshape existing audio; /repaint takes fractional seconds |
/stems | Split a mix into vocals and instrumental |
/split | Split a mix into vocals, drums, bass and everything else |
/score | Reads a stretch into a score: bars with their start times, chords, sections, key, tempo and (optionally) the sung words |
/lyrics | The sung words as plain lines, written down by HeartTranscriptor from the separated vocal |
/restage | Covers a stretch with YuE2: the same melody and chords, sung or on an instrument, locked to the original bars |
/band_score | A YuE2 band that plays the take's own chords and form, locked to its bars, under the real voice |
/score, /restage and /band_score work through the score rather than the sound.
SheetSage2 reads the melody, chords, sections and every beat of a stretch; scores.py edits
that score (chords per bar, transposition, tempo, giving the vocal line to an instrument,
removing it for a band) and rewrites it in the native two-voice ABC, checked with the official
abc_tools.py parser (vendored, MIT). YuE2 then plays the score, so the tune is the original's.
YuE2 has no timing or audio-reference input, so sync.py reads its render again, pairs its bars
with the score's bars by harmony (dynamic time warping, tolerant of a bar added or dropped), and
Rubber Band stretches it until every bar starts where the recording's bar starts. A new tempo
disables that lock, since the original bar lines no longer apply.
No model re-sings you. /accompany writes a band and mixes it underneath your audio.
A tempo and key in the prompt are only a hint to ACE-Step, so a band written from text alone starts on its own downbeat, drifts and plays its own chords. It follows audio closely, though. So the engine:
separate="auto"). If there is some, the voice
is lifted off, and the old backing is what gets analysed: it carries the beat and harmony far
more reliably than a melody does.strength sets how closely the band keeps the guide's timing and chords.split_parts it also
returns drums, bass and the rest as separate stems.It returns the mix, the band, your vocal, the prompt, the original backing (when one was found), the three band stems and a JSON summary of the tempo, key, beats and chords it used.
Singing2Accompaniment is still unreleased, so the band follows your beat and chords, not your
phrasing. Rubato takes keep their tracked beats instead of a fitted grid, but unmetered singing
with no clear pulse falls back to text-to-music at the measured key.
python test_band_tools.py checks the analysis on synthetic audio with a known tempo and
progression; python test_audio_tools.py checks the mixing.
They pin conflicting versions — transformers 4.45.2 (SheetSage2), 4.50.0 (ACE-Step) and 4.57.6
(YuE2) — so pip cannot satisfy all three. The YuE2 wheel's stack is installed and the other two
run on it; ACE-Step is installed with --no-deps and its inference dependencies are listed
explicitly in requirements.txt.
Each model is built on first use, not at import. YuE2 requires this: it reaches low-level CUDA
init while constructing, which ZeroGPU refuses outside a @spaces.GPU function. Weights are
fetched at startup, so no request pays to download them, but the first call to each model
carries its load time and its GPU budget allows for that.
Requests are queued one at a time, uploads are capped at 100 MB, and temporary results expire after 24 hours. Call with the signed-in user's Hugging Face token to use that user's ZeroGPU allowance.
SheetSage2 and YuE2 are CC BY-NC 4.0 (non-commercial). ACE-Step is Apache 2.0 and Demucs is MIT,
so /accompany, /part, /instrumental, /stems and /split results carry no non-commercial restriction.