tested BPM conditioning and key/harmonic conditioning using ABC.
Hi everyone! I recently started testing YuE2-3B more systematically and wanted to share a few observations from my experiments.
So far, I have mainly tested BPM conditioning and key/harmonic conditioning using ABC.
For BPM, my tests suggest that the BPM value acts more like a target/conditioning signal than a strict tempo lock. The generated audio can follow the intended pulse, but perceived tempo may still vary depending on the musical context, including half-time or double-time interpretations.
For key/harmony, I tested different K: values and harmonic anchor combinations in ABC. The results suggest that the model does respond to the supplied harmonic information, but it does not behave like a rigid MIDI sequencer. With fewer anchors, the harmonic relationship tends to be more clearly preserved; with more anchors, the model may reinterpret, reorder, or select parts of the harmonic material while still producing coherent music.
I’m continuing to test these behaviors and will share more results as I get them.
Just posting this in case these observations are useful to other people experimenting with YuE2-3B.