Thanks
Glad to see these available in a pretty easy to inject into KoboldCPP.
Though the styles (as of Ace Step 5/6 Turbo) are a bit more limited not doing heavy metal, rock, tribal or others very well. And often outputs sounding like it was recorded using a Yamaha keyboard from the 90's with a 2Mb Soundfont chip. I'm sure these will improve as newer models are put out.
Even putting the largest models with 8Gb Vram get decent output in a reasonable time (like 5-10 minutes per song), probably due to low VRAM option which swapping models as it needs rather than trying to hold the whole thing at once or relying as much on CPU.
Glad you like it!
Trying the newer Yue2 with 1.124, and while it failed to generate lyrics (wait... i don't see a 'plan lyrics' button)... anyways, the nonsense that came out is actually quite catchy...
Weirdness aside...
@yano2mch You need to load an extra text LLM to generate lyrics. I recommend a small lightweight model like this
https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct-GGUF/resolve/main/Qwen3VL-4B-Instruct-Q4_K_M.gguf
Place the lyrics into the "Lyrics" textarea box in MusicUI http://localhost:5001/musicui
Alternatively, you can open up the normal koboldcpp and launch your own text model, and ask it to make some lyrics for you.
@yano2mch You need to load an extra text LLM to generate lyrics.
Mhmm, so i noticed. Though in this case when i did have another model loaded it just took my suggestion/idea prompt and copied the whole thing verbatim as verse 1 and then nothing else.
I just thought the lack of lyrics and getting the resulting half mumbling words garbage as a pseudo language was.... interesting... GF says some artists sing this way so it could totally be on the radio and not be noticed.
when i did have another model loaded it just took my suggestion/idea prompt and copied the whole thing verbatim as verse 1 and then nothing else.
you'd put the prompt idea in the "Music Description" box, leave Lyrics empty, then hit Plan Score. I recommend a small Qwen model for this.
Output (note that I used the SheetSage2 notes in ABC score for a certain popular song here, in melody cover mode)
SheetSage2 is coming to Kobold in the next patch release. Stay tuned.
Patch release is out
when i did have another model loaded it just took my suggestion/idea prompt and copied the whole thing verbatim as verse 1 and then nothing else.
@yano2mch
you'd put the prompt idea in the "Music Description" box, leave Lyrics empty, then hit Plan Score. I recommend a small Qwen model for this.
Mhmm. I remember trying that in previous versions; Resulting lyrics were a bit weak (not your fault i know). Just had half expected Yue2 would have generated lyrics.
I'd verified alrady adding a 'main LLM' for it to generate lyrics from and it worked, assuming you had another LLM loaded. The description i'd used for the above pseudo-language song was probably just weak enough it just copied it. Again not your fault.
Doesn't change that i appreciate your efforts :)