Hey guys, thank you so much if you bought the software, as you bring me the $$ and also the motivation to keep working on this.

Playback:

Playback is now overhauled and it works off of a time based system (which to be honest is what i should've started in the first place) so this means that playback will not speed up when generating or any of that nonsense

NEW MODEL UPDATES

(Note: you must require an internet connection as the new models are hosted on the internet, the model files should cache though once you boot up the kobimusic model file)

The OpenMusenet-2.1 models have been updated to OpenMusenet-2.11, which adds a missing feature in transcribing the midi dataset, the note lengths not being scaled. It is a bit complicated to explain but lets say a midi note is being played for 48 units. The definition for units to quarter note values is defined in the header of the midi, which is very often different for different midis. This oversight I think decreased model performance in rhythm likely a significant amount and so now, the ai scales those 48 units to the same definition for all midi files. This also made midi files be played at pretty much a random speed (did not correspond to the actual midi bpm if you are wondering)

Also I added 1 new serious model and also one absurdly small model: OpenMusenet-2.11-S and OpenMusenet-2.11-3M. OpenMusenet is about ~35% less size than the base model and OpenMusenet-2.11-3M is.... 45 TIMES less (Note this doesn't mean it will run 45 times faster sadly due to encoding tokens and stuff).  The model itself is literally 14mb which is literally microscopic by even AI standards 5 years ago.  Also, the kind of weird thing about the OpenMusenet-2.11-3M model is that it isn't probably as horrible as you think, it kind-of vaguely understands melodic structure and can generate simple chords.

Also as a small bonus, I added an early version of OpenMusenet-2.11 (large), trained on video game music if you are interested...

Also here are the model capabilities.

MODEL CAPIBILITIES:

OpenMusenet-2.11-L - Most of songs generated from scratch are arguably "good".  Understands rhythm to decent degree. Can be Really good/creative at chord progressions. Can typically continue style from complex user-made song in 1-2 "generate" clicks

OpenMusenet-2.11-M - Similar in performance; bit worse, for generating from scratch. At least basic concept of rhythm. Decent at chord progressions. Struggles some more than the large model, but typically eventually gets a decent result from user-made song.

OpenMusenet-2.11-S - Sometimes decent from scratch, sometimes not. Basic concept of rhythm.  Can understand chord progressions to a lesser degree of the medium model. Strangely good at continuing some semi-complex songs but struggles on more complex tasks.

OpenMusenet-2.11-3M - Occasionally decent strangely. Spams lines of 8ths or quarter notes, sometimes even both. Writes only one chord most of the time, typically coherent chord. Do NOT try prompting this with anything more complicated than hot-cross buns, it will LITERALLY throw a fit and mash the adjacent notes, try for yourself lmao.

OpenMusenet-2.11-VG - Similar in capabilities as the normal large model, but not as good at continuing user-made music.

Anyways, that took way long to write but yeah, here ya go!