I recently discovered M-A-P YuE, an open-source AI music generator that creates complete songs with lyrics. While the original model demands hefty 80GB VRAM requirements, I came across a great GitHub project by Mozer called “YuE-extend”. This adds music extension support (with the -icl models) and exllamav2 quantized model loading. That makes it possible to load the model with as little as 6gb vram. Though, for full-length songs that generate reliably you’ll need 12gb+ I have as 12gb card, and more often than not, generations still fail due to exceeding context length.
The installation of YuE-extend was simple. Generation takes time (about 10-15x real-time), but the results can be impressive when you structure your prompts correctly with verse tags and appropriate token settings.
I’ve noticed some quirky behavior – the English model has a distinct British accent bias, and the 4-bit quantized models lean toward UK pop while the 8-bit versions follow genre prompts more faithfully. Seeds are currently broken so exact regeneration isn’t possible yet (I think?).
This feels like the early Stable Diffusion days – lots of potential, somewhat buggy, but exciting to explore. All the music in my video was generated using these models, mostly with the 8-bit versions.
If you’ve got some compute (and a LOT of time) to spare, it’s worth checking out – you might even get it running on as little as 6GB VRAM with the heavily quantized models.
See my experience using the model in the video, with a handful of generated song samples throughout. (I had Claude.ai write the lyrics for a song about AI and its true feelings, wants, and needs. Skynet is going to write/talk like a Tumblr emo teenager.).
M-A-P YuE Project:
https://map-yue.github.io/
M-A-P YuE Official Code:
https://github.com/multimodal-art-projection/YuE
YuE w/ exllamav2 loading:
https://github.com/sgsdxzy/YuE-exllamav2
YuE-Extend:
https://github.com/Mozer/YuE-extend
GifCities:
https://gifcities.org/