KobiMusic
A downloadable AI music generator for Windows
KobiMusic
KobiMusic is an easy to use GUI for our state-of-the-art MIDI generation AI. You can prompt it with your own/others' music and can get a coherent completion to it. We also offer the music generated by the AI under an MIT licence meaning you can sell the music generated by the AI with no attribution!
Why should I buy?
This can pretty accurately continue or create new songs. You can also input the genre into the AI. The quality I think atleast is much much better than any of the AIs that work off of waveform. ALSO there are pretty much no alternatives for MIDI generation (MuseNet was shutdown and Magenta AI is pretty lacking...)
Demo
Before buying the software, I very highly recommend running the demo first incase the build does not work on your pc. It is also a great test for the AI's quality or if your just on the fence about buying. The demo does not allow saving a midi and is limited to 3 minutes per session.
Samples
Note that the actual music, parts and stuff like that were solely and entirely by the AI, but some of the instruments were changed around and it was recorded in a DAW
From scratch
Dire Dire docks reimagined
An original Jazz compisition
You can also look at the KobiMusic youtube channel for more music
What is a MIDI file?
A MIDI file is a file that contains musical instructions rather than actual audio, like a set of sheet music for electronic devices. Instead of storing the sound itself, a MIDI file stores information about which notes to play, their duration, and other musical parameters like volume and pitch. This allows MIDI files to be easily edited, played back, or even used to control different instruments or software.
Bugs
If you can't launch the program, send me a comment somewhere and I will try to fix the build ASAP
How do I use?
Here is a vid explaining how to use the program
System requirements
Windows 10+
Decent GPU is a nice-to-have for larger models
Large model needs 8GB of ram.
Actual app info
The app itself is pretty simple, Prompt the AI with specific genres using the Guide AI button. A list of genres that work well can be found here, just paste in the text from the genres column. You can open/save in the top left, player controls in the bottom left. There is also a help in the bottom left as well since I don't feel like explaining all the controls. Drag around your view with left cursor and set player position with right mouse button. You can remove notes to the right of the red line with the R key (no undo so be careful!) and yeah. Like I said check out the in app help for more info
Speed
Each generation takes ~25 seconds on my 1080, this also runs without gpu if you have none (or if the app can't detect one), and takes probably 2 minutes. On one of my blog posts, i even got it to run on a school laptop if you are interested.
Purchase
In order to download this AI music generator you must purchase it at or above the minimum price of $14.99 USD. You will get access to the following files:
Development log
- Midi saving hotfixes57 days ago
- A new 2x CTX model61 days ago
- A list of genres that work well with the new Maestro-3.11 modelDec 03, 2023
- (finally) Better prompting update!Dec 03, 2023
- Going on cruise tomorrowOct 12, 2023
- Basic framework for guiding AIOct 04, 2023
- General Improvements update!Sep 29, 2023
- Update on upcoming update...Sep 28, 2023
Comments
Log in with itch.io to leave a comment.
i love it and i would get it, but its paid :(
Sorry... :(
for some reason, most of the midis i try to load in just, dont load at all. i dont know why this happens, and it doesnt display an error when it doesnt load, and honestly i dont know what to do.
This is due to the strict error handling in the midi loading section of the program, but for now you can boot https://signal.vercel.app/edit and just load the midi in the file dropdown, and just save it then reimport the midi into the program
I might be in over my head..
I'm getting an error that reads
File "transformers\models\gpt2\modeling_gpt2.py", line 322, in forward
torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 20.00 MiB (GPU 0; 4.00 GiB total capacity; 2.56 GiB already allocated; 10.64 MiB free; 2.61 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
Any ideas?
This error means your GPU ran out of vram when trying to load the model files, so try closing out of any other apps (if you have any other open), and you may want to try restarting your computer too for good measure, then rebooting the software.
Generation works fine, but whenever I save what I got, it saves as a 0 kb midi file that has no data in it. This happens every time I try to save anything I've produced. What's going on?
The console also gives this error, which I assume is what's causing this.
EDIT: It seems like this happens only if I do too many generations at once, like more than 4. It seems I'll need to save the progress frequently or I may lose all my generation work.
EDIT 2: No, it seems that it can break even with one generation. I am LUCKY to be able to save anything I make in KobiMusic...
I also have the same issue MayroSMM was having, though the workaround site you linked did work for that at least.
I dunno man. I get you're working on this solo and trying to recoup some losses, so I'm not trying to be mean here. But while these generation models you put together are cool and worth paying some money for, the program itself is very basic and buggy to the point of losing work and needing to restart it to restore function. If this cost like $5 I wouldn't really mind, but at $15, I'm kinda feeling like I spent too much for what I received. I expected something a bit more polished at that price tag.
Hopefully I can get some support. >_<
Bruhhhh how did i mess up this badd
I am terribly sorry for all of this experience and lemme refund you rq (Itch.io must process it so it will likely be a few days)
I'm gonna try and debug this, i suspect i know the problem (it is an issue in the new update) but it may be upwards of a few hours.
Aw, well I appreciate that man! :O I don't REQUIRE a refund, I'd just like a functioning program more than anything, but if you feel like doing that it's your choice. I admit I've gotten very frustrated today in failing to use it, but if the bugs can be solved that would be great. I think your program has potential, it's just not there yet. (Kinda feels like it needs a $5 Early Access version or something until it's all polished up and maybe expanded a bit, then $15 would be fair. I guess Itch doesn't really have an Early Access feature though.)
Here, let me give you another console log I noticed. When I start the program up, it actually gives me this odd message which might relate to the other problem? Perhaps this will give you more to work off of.
The 'fluidsynth error' line happens at the end of program startup. It's very confusing as I have a perfectly-functional MIDI device on my system.
The stuff following that fluidsynth error line happens when I try to do my first generation. (The generations do seem to work and play fine, but it seems something is buggy with that process; the message certainly sounds ominous at least.)
Oh, managed to trigger one more error. Not sure how, but delivering.
Okay, the new build is up now, so hopefully upon redownloading it should all be fixed 🤞
Nevertheless, lemme try to get itch.io to refund you for this bad experience
Appreciate it man. I gave the new build a run. Unfortunately, now this happens after the 'login successful' prompt, and then everything immediately closes an instant later.
Apologies again, i seem to have mistakenly forgotten to include fluidsynth in the build
New build up again, hopefully should work this time 🤞
What prompt is the best for generating video game music? Thanks for the great tool!
I would say it really just depends on what the purpose is, so like maybe for a boss battle theme you might want to try something like "metal" or "dance-eletric" (note the spelling), or like for an overworld theme, you might want to try something like "film", maybe "soundtrack"
Got ya, thanks!
how to extend songs with the ai? when i generate it doesnt make anything over 3 seconds.
You can press generate again after generating a segment you like, and it will continue from that new segment and the previous. Right now it is indeed limitted to generating in small chunks
alright thanks
Any plans for a Mac release?
Yes, but probably not for a while because getting the app to compile to mac-compatible format has proven really difficult.
Thank you!
I gonna spread 50k on this
huh
lol wot
Hey, when I export the MIDI it puts everything onto one track rather than keeping the original tracks. How can I prevent this? Also, it would be awesome to have a feature that generates a certain length.
The tracks are color coded. When you import your midi, is it all one color? If it is, i would recommend going into an app like signal.vercel.app, then try removing some extraneous tracks if there are any and then update the channels to be just in the first four (since the app separates just the first four channels, then sets the rest to one channel).
Generating off of a certain length would be a good idea, when i'm able to create larger context models (For now though, it is probably not worth it because the current ai models really needs the context)
When I import the MIDI, they are color coded yeah. But when I export, it exports as just one channel/color. I'll try the app you suggested and get back to you anyway :)
Hmm... are you trying to re-import the midi into the program by any chance? You may also need to also resort the channels in an app as with importing.
This is kinda buggy behavior so im gonna go ahead and add this to my bug list that im fixing and will fix this soon
I tried putting the MIDI back into kobimusic today out of curiosity and it actually does recognise all the channels in there. I was specifically using a Super Mario Bros. midi, as I wanted to generate new sections of the OST, drag the midi into gxscc to emulate the 8-bit sound without having to actually worry about real limitations of the NES, and then blind react. But it puts everything onto one piano channel. Drums and the lot of it. Same thing when I tried loading it in SynthFont, same thing when I open it with Windows Media Player, and same thing when I sent it to a friend. One thing I noticed is when I go to save the midi, it does not have a recommended file. It's not prompting me to name a .mid file, it's just blank and you gotta add the filename yourself. So maybe something in the saving is getting mixed up?
Ah ok, I was able to replicate this bug with the gxscc app. I was able to fix it by importing it in this, then just simply downloading it again. You can also configure the instruments there if the gxscc app doesn't support changing of instruments.
Hope this works for you🤞
can i please get a free trial please.
na sorry bro just download the demo tho if you just wanna try it out
ok
When I click the button to generate, it continues to say it's generating, and the percentage still goes beyond 100% (like 200%), and the time turns negative. Any fixes?
The app also eats up my CPU usage.
My Laptop uses AMD by the way, I am not sure if that changes anything.
Does your laptop have a GPU by chance? Generation should still work (albeit pretty slow) if you just have a CPU. The time/percent goes over 100% since it is just an estimate of the time remaining.
yeah but I switched to my other laptop which uses an intel cpu and a old gpu and that is endlessly generating also
it got up to 936432% done before I shut it off (what the heck by the way)
the app just never stopped "generating"
Ah, ok, you might be running out of ram then, mb for the inconvience bro. That means that you are probably limited to using the smaller sized models.
Sorry bro, lemme get you with a refund rq (through itch.io so it might take a while)
nah it worked after I switched the model, it generates a very small amount at a time but it works after like 30 seconds per one section now. Idk what was going on before.
Just to ask though. OpenCL? Probably not possible/easy to do.
Was just wondering, please don't go crazy on trying to make it a thing.
It actually does seem possible to do with pytorch according to this guys implementation but going off the benchmarks, it looks slower.
Doesn't do anything still on midi 1 files.
Possibly because on cpu. But it goes to 59857%+ and nothing. So. Yeah, idk man.
That is very strange and idk man. I'll refund you for that as that is obviously a big issue and im sorry for the inconvenience. It should still work on cpu. Can you detail me on the specs of your pc and were you able to generate in the demo version ?
(Note, the refund may take atleast some time because i have to contact itchio support)
Hey man. Don't worry about refunding. I'd like to support your work.
It did generate eventually, not sure how consistent it is. As things are just crazy, but to me, it's worth it. If worst comes to worst, I'll hustle up a gpu, don't worry kin_g
just to inform you though, I was using my i5 1145G7 on laptop. When it did generate, it took about 45 seconds to complete 120%. Where it actually posted the DONE and showed generation.
There are about 5 older laptops/pcs I could probably use to test cpu otherwise.
I would not have purchased without confidence :)
A simple how to video would be very much appreciated. As it is loading MIDI files is hit or miss. Clicking generate doesn't seem to change the music at all. Pretty sure most of it stems from my lack of knowledge on how to use the software.
Hmm.. Ok, I will make a video. One sec
Alright heres the vid
Hope this helps 🤞Okay so I think I know the source of your midi import bug, as it turns out, the library I am using for reading the midi files will completely fail on midi 0 files as opposed to midi 1.0 files which work fine. This may take a day or two to fix 😅
Ok, after a little while of searching, there seems to be no easy way to do it in the programming language i am using to make this program. I did find an external program that will do it for you though if it helps a little bit http://downloads.fyxm.net/GN1:0-MIDI-66164.html or, for a web version, you can goto signal.vercel.app then import your midi and download it again.
MIDI file type convert · GitHub
Yodi, is pay to use.
I really hope the midi import can get sorted soon. I've been enjoying using it though a bit annoyed my Sekaiju, OpenMPT or Logic Pro exports don't pick up instruments properly, I have to export it via FL Studio