Good question. First and foremost, Magenta is autoregressive, and this is not. Practical differences:
THIS:
- has more controls; output can be highly custom
- has very fast response time for updating controls; this approaches real time
- runs on consumer cards
- is optimized for loops, and full songs, not continuous infinite music
MAGENTA:
- can generate music forever, it plainly wins here
- has fewer controls
- has high parameter update latency (chunks of 2 seconds)
- I did get magenta to fit on a 5090, but just barely
what was the biggest unexpected unlock you got while building this? like was there something that just worked out unexpectedly easy compared to the rest?
also, another question, how much time it took you to build this thing all together?
biggest unexpected unlock i would have to say is the shared mutable curves that bypass the ringbuffer drain. i wouldnt have thought of that without the throughput that preceded it
Wow this is so cool! Do we have VST support?
can you explain performance differences vs Magenta / Lyria? that's the most similar thing I can think of
Good question. First and foremost, Magenta is autoregressive, and this is not. Practical differences: THIS: - has more controls; output can be highly custom - has very fast response time for updating controls; this approaches real time - runs on consumer cards - is optimized for loops, and full songs, not continuous infinite music
MAGENTA: - can generate music forever, it plainly wins here - has fewer controls - has high parameter update latency (chunks of 2 seconds) - I did get magenta to fit on a 5090, but just barely
Let's talk more about why it is reason number 2.
lets talk philosophy, bro
incredible piece of work
thank you kindly
this is sick and why not bro!!
what was the biggest unexpected unlock you got while building this? like was there something that just worked out unexpectedly easy compared to the rest?
also, another question, how much time it took you to build this thing all together?
biggest unexpected unlock i would have to say is the shared mutable curves that bypass the ringbuffer drain. i wouldnt have thought of that without the throughput that preceded it
oh and as for time - this is several months of work, but is the culmination of several years of audio adjacent genai work