ACE-Step 1.5 behind Kiku AI, built by Reubencf. Its endpoints take the same arguments and return the same outputs as Kiku-Engine, so the studio can send a request to either model.
| Endpoint | ACE-Step 1.5 task |
|---|---|
/song | text-to-music with lyrics, planned by the 1.7B 5 Hz language model |
/instrumental | text-to-music |
/remix | cover (restyle), strength = how closely to keep the source |
/repaint /extend | repaint, inside the audio or past its end |
/part | lego: one instrument written into a mix (base model) |
/accompany | complete: a band written around a vocal (base model) |
It is its own Space because ACE-Step 1.5 needs Python 3.11, Gradio 6 and torch 2.9, which do not fit Kiku-Engine's YuE2 stack.
/accompany never keeps a voice the model wrote. Demucs first lifts the voice off any music already
on the take. The base model's complete task then writes drums, bass, keyboard and guitar around
the vocal, conditioned on the audio itself rather than only on a tempo in the prompt. Demucs
removes whatever vocal the model rendered, and the band is mixed under the original take.
The acestep package is vendored from the official ACE-Step 1.5 Space (MIT, see
LICENSE.acestep), without its UI, training and nano-vllm code. Results may be used commercially.
ACE-Step 1.5 behind Kiku AI, built by Reubencf. Its endpoints take the same arguments and return the same outputs as Kiku-Engine, so the studio can send a request to either model.
| Endpoint | ACE-Step 1.5 task |
|---|---|
/song | text-to-music with lyrics, planned by the 1.7B 5 Hz language model |
/instrumental | text-to-music |
/remix | cover (restyle), strength = how closely to keep the source |
/repaint /extend | repaint, inside the audio or past its end |
/part | lego: one instrument written into a mix (base model) |
/accompany | complete: a band written around a vocal (base model) |
It is its own Space because ACE-Step 1.5 needs Python 3.11, Gradio 6 and torch 2.9, which do not fit Kiku-Engine's YuE2 stack.
/accompany never keeps a voice the model wrote. Demucs first lifts the voice off any music already
on the take. The base model's complete task then writes drums, bass, keyboard and guitar around
the vocal, conditioned on the audio itself rather than only on a tempo in the prompt. Demucs
removes whatever vocal the model rendered, and the band is mixed under the original take.
The acestep package is vendored from the official ACE-Step 1.5 Space (MIT, see
LICENSE.acestep), without its UI, training and nano-vllm code. Results may be used commercially.
