Source on GitLab.
Why this exists
I wanted to play around with automated music generation, and I wanted it to actually be about something rather than just vibes-and-prompts. The obvious answer was to make it research its own subject, write real lyrics, and then inflict those lyrics on a music model.
It’s called stupid because that’s the honest description. It is not trying to be impressive. It is trying to make me laugh.
What it does
The pipeline goes: crawl the web for information about a given subject → feed that into a RAG context → run an agentic loop to generate and refine lyrics within whatever constraints you give it → hand the result to either Suno (manually, for now) or ACE Step 1.5 for audio generation → render a visualiser video with lyrics synced to it.
Output is either a 60-second clip formatted for Loops.video or the full track with video. The whole thing is mostly parameter-driven — tweak the constraints, the subject, the genre framing, and see what falls out. It seems to do particularly well with psychedelic rock, for reasons I have not investigated.
Current state
It runs. The Suno path works fine — it’s just manual, which is annoying but liveable. The ACE Step 1.5 path is currently blocked: the LLM side wants to run ACE in vLLM, and vLLM dropped Pascal support. Since I’m running Pascal hardware, that’s a wall.
The fix is integrating pascal-attn so ACE Step can actually run. Once that’s done, the full pipeline closes and ACE is fast enough that this becomes genuinely usable as a toy.