วีดีโอ

Thanks Dirk. I've still got it saved in Udio and maybe I'll try lengthening it, see what it does.

With this piece, I got the same feeling listening to the output for the first time as I had with some of my best MuseNet pieces, even though with those, of course, the composition process was more collaborative. And some of the other things I generated with LoudMe are pretty nice, too, although not like this, unfortunately. But like this, I will need to convert it to MIDI to work on them, and also to just obtain a better audio quality. As I mentioned in my comments below the video, I blew through all my free credits to use LoudMe on the first day of using it. It let me generate about 20 pieces before I hit my free limit, and the pieces are mostly between 2 and 4 minutes long. Udio, however, let's you work by training their model, to generate 10 free credits every 24 hours, and not only gives you 100 new credits a month, but also gives you 10 credits every day. If you don't use it for two days, you will not get 20 credits though. But later in the day, before I had used all my free credits with LoudMe, I think maybe I did something to increase my chances of it not putting vocals in the music even when I have "Instrumental" selected. Instead of starting the advanced interface by letting it "auto generate," I selected "instrumental" first and in the top lyrics box I put only the letter "a". I wish I had more free credits to experiment with what prompts do what. Before I subscribe to LoudMe, start paying the $8/month to use it more, I'd be interested in hearing the opinions of my listeners, what your experiences were like with Suno, Udio or LoudMe. Just recently, Udio let me produce a classical sounding vocal where the voice is excellent, and the piano quality satisfactory, and LoudMe won't do that. It has a kind of grit in the sound. However, considering the 20 generations I did with LoudMe, the musical and structural qualities of the pieces are more consistently good than those I've done in my few weeks on Udio. If anyone didn't read the comments below this, I spent hours fixing this piece in my MIDI software, as the audio-to-MIDI conversion I use is rarely perfect. People who are less inclined to go through what I did with this, removing the vocal spots, doing a conversion to MIDI, they'd be better off trying Udio.

​@@Timzart7 As I mentioned in my earlier comment, LoudMe is using Suno to generate its music. I uploaded a couple of videos on my channel over the weekend showing this. I would be very cautious about paying for LoudMe. I use both Suno and Udio (and am one of the "guides" on Suno's discord server.) I find Suno gives pretty nice song structures, and works well for punchy/short instruments (plucked strings, percussion etc.) It generates up to 4 minutes to start with, but my personal preference is to keep extending from an earlier point as I like to change things up as the music progresses. Udio has really nice vocals, seems good for smoother sounds (bowed strings, synth pads, etc.) and has a wider sound overall. I haven't tried making percussion with it recently but it at least used to sound a bit watery for sharper sounds compared to Suno. The shorter generation length doesn't bother me much but I do like to be able to start listening almost straight away with Suno. Both have upload functionality so I sometimes take audio from one and upload it into the other which can help to refine the sound quality a little. I'm mainly interested in getting samples from them to build upon.

@@FurrybeansMusic Good to know, thanks. I watched both your videos on it at your suggestion. I read another caution about paying for LoudMe. And since I only used Suno for less than an hour, I'm going to have to try that more, especially if it is, as you suggest, as good as Loudly, because Loudly is using Suno. If I were not interested in mainly piano music, I'd be impressed with the sound quality of SOME of the Udio output, like a classical song I generated that sounds like Schubert, and the voice is beautiful. LoudMe, to me, all the output sounds fuzzy, or slightly buzzy in places. And the sound quality of the Udio French Pop piece I posted, "Sous le ciel de Paris," is quite nice in its raw state. I'm trying to generate piano things that sound like either Debussy and Ravel, and since these generators balk at suggestions you generate some music that sounds "like" a specific composer, and Udio rarely produces something when I put "Impressionism" in the prompt, that sounds even vaguely like that style, which is use of the whole tone scale combined with chromaticism, I guess I'm going to have to be more creative in my prompts. If you have any suggestions on the best free audio-to-midi converter, I'd be interested in trying some others. I just put what I thought was a simple Mozart-style piece into the one I use this morning and it really demolished it. It was too staccato or something, and it made the converter throw the rhythm off. I'm going to try quantizing it.

I've composed over 200 AI pieces,, starting over two years ago. I try all of the available AI music generators that I read or hear about. I did try Suno but don't remember what I didn't like about it, probably like Udio, Suno outputs only in audio formats. I like the collaborative generators more, like MuseNet, my first, but that stopped working and so I've been using mostly Python programs on GitHub, which also require more time and choices by the user. Two things brought me back to Udio a few weeks ago, after initially rejecting it, because it doesn't output in MIDI. First, I heard a few compositions that I really liked. Then, Alex, the same guy who writes the Python programs I use, also wrote an audio-to-midi converter for piano that works pretty well. So, I can output on Udio in audio, and convert to MIDI and edit piano pieces. I'm interested in that process, but at the same time, the way Udio sounds, and some of the things it produces, I can't ignore it because it is foreshadowing what is coming in AI music composition -- great things. One piece of advice on using Udio is that if you can generate in the off-hours, it generates clips four times as fast. Glad you like this piece.

So, when doing these sequences using a transformer/sequencer, you can "continue" a seed MIDI, measure by measure, or many measures at a time, or you can "inpaint" it. Inpainting can do a much longer piece/seed, and what it does is it transforms the notes in the entire seed, measure by measure, even if the seed is over a minute in duration. With continuations, it is better to use a seed file of only 10 to 30 seconds long. Inpainting is also fast to do, in Allegro. Sometimes I inpaint just a short piece, or sections from a piece, and sometimes I do try to inpaint an entire piece although with long pieces, it is better to cut them up into about one minute segments before inpainting. To give an example, let's say you input Debussy's Clair de Lune as a seed file, although you'd need to cut that in half probably, as it is too long. So you load that in as a seed, in MIDI, and then it inpaints the whole thing, and does just kind of a variation on each measure. But to experiment with inpainting the first time, just load maybe 30 seconds of music (in MIDI format), and do piano to keep it basic, and then after that is loaded, go down to the inpaint cell and set the parameters. Then click on the arrow to start the inpainting. When it is done, the file will appear in the file menu at left as "Allegro composition.mid" or something like that, with no number. The ones with the number are the continuations. The one slider/parameter to pay attention is you get to choose how many tokens you want the inpaint to borrow from the original. I'd recommend 16 to 30 or so. Those notes will be the same as the seed when you play it back. You'll see what it does after you try it a few times.

Two more things I should add. The Inpainting feature is in the "Original Version" of Allegro Music Transformer, not the "Composer Version." And the easiest way to try inpainting is to use the Hugging face app I will link below for inpainting. You just drop the midi file in the space, and click on inpaint, and then it will let you play the MIDI and you can download the inpainted MIDI if you like it, by clicking on the MIDI cell, little text that shows the size of the MIDI. This is a much easier way to do an inpaint than the GitHub version of Allegro Music Transformer: huggingface.co/spaces/asigalov61/Inpaint-Music-Transformer

​@@Timzart7 Oh, okay! Thank you so much for the explanation, very detailed. I'll try it myself right now. By the way, what do you think is the best model by Alex for solo piano pieces? I've seen that he also has the QUAD model now, and there's also the GIGA-PIANO-XL, but it’s older. I'm also checking out the Giant Music Transformer that you recommended.

@@manushi27 I do mainly piano, too. Alex's GIGAPIANO is a little glitchy and dated. I tried it myself. The QUAD model is good for choral music. It is very fast. Actually, today, I was using Allegro Music Transformer, Original Version, just for some continuations, even though Giant Music Transformer Bulk Generator is very fast and convenient. I just haven't had good luck with the quality of Giant's continuations and inpainting lately. I'm also fond of Allegro Music Transformer, HuggingFace version for seed generation, one at a time. It is slower, but easy to just go there and do it, without the fuss of loading all those cells in GitHub. Here is a link to the Seed generator: huggingface.co/spaces/asigalov61/Allegro-Music-Transformer And just this week, I'm using Udio more, which is a remarkable leap forward in the generation of AI audio files. I text prompt with "piano" in the prompt, hoping it will generate a piano piece, and then because for free users, it outputs in MP3, I usually drop that Mp3 in Alex's ByteDance conversion HuggingFace app that converts an Mp3 or Wave into a MIDI and then I can do whatever I want with the piece in MIDI. And Alex's app is for piano audio files (wave or Mp3) only. Here's a link to it: huggingface.co/spaces/asigalov61/ByteDance-Solo-Piano-Audio-to-MIDI-Transcription

Thanks Elly. I'm going through a dry period in creating new AI pieces but I'll get through it. I thought of you when I listened to this guy's arrangement of the Sorcerer's Apprentice for clarinet octet. But can you believe it is software playing it? It sounds so real! Here is the link: th-cam.com/video/VoCoGSl_SIM/w-d-xo.html

@@Timzart7 dry periods come and go..but the talent remains!

It is there, but doesn't work. In the "advanced mode," click on the arrow at the right to generate new measures. Nothing happens. MuseNet is dead. MuseNet stopped working on December 12, 2022, just one month after ChatGPT made its debut. ChatGPT, which shook the world, earns tens of millions of dollars for OpenAI and in comparison, MuseNet was more a vanity project, used by a few thousand people over the three years it worked, starting in 2019.

Glad you found it helpful! It's amazing that any tech tutorial that is 12 years old still works. I guess there's a certain continuity between versions of Windows. My last 200 posts on TH-cam are my music compositions made with various AI music tools, so you might want to unsubscribe. I did make a few tutorials on the AI music tools though.

MuseNet, which stopped working in 2022, featured 12 styles to choose from, and I enjoyed using the Tchaikovsky, Chopin and Rachmaninoff styles especially. However, after a thousand hours working with AI composition, I moved toward working with AI composition software that allows me to work in any style. Here, for example, is a piece that was done using a "seed" file, consisting of maybe 8 measures, from Debussy's Reverie, and I think it is pretty successful at sounding like Debussy, without it sounding much like Reverie at all: th-cam.com/video/y53CjztrF0Q/w-d-xo.html Lately, I've branched out into more modern styles, exploring the possibilities for a new style more related to the nature of the sequence characteristics that AI generates on its own. Here is an example of that: th-cam.com/video/pd0yt1DbfJM/w-d-xo.html

Haven't tried Udio, but have tried several other audio output apps, including the ones touted as the most recent and best. I don't like distortion in the music, and most of the audio-output-only apps have some degree of it. I'm sure it will get better over time. But the main thing is, Giant Music Transformer is collaborative, and for composers/musicians who, like myself, have worked with MIDI for years composing and/or arranging music. I like being part of the creative process with complete control over every note and phrase.

@@Timzart7 also, there is a growing library of user compositions, it is all public and is slowly becoming a great platform even for casual listeners

MuseNet died on Dec. 12, 2022. OpenAI discontinued it. What you now see on the MuseNet page is just a demo, not the usable advanced interface it once had. The sign-in on the MuseNet page is for ChatGPT or API use only, not MuseNet. When it did work, MuseNet had no sign in.

Thanks. I did exactly that, used the piano roll view for the video on one video (of my 200+ AI pieces) anyway. No one commented they like it so I never bothered doing it again. And anyway, this video is getting a handful of views and not one "like" which would attract a few more views. Gone are the days when I got 1000 views in one day on an AI piece. That just happened on two pieces and was a fluke. I should cross-promote on other platforms or forums, but I'd rather work on music, because as you know, building a piece and refining a performance takes a lot of time.

Being retired and having lots of time, I've done music composition using AI sequencers for coming up on two years now. I've posted over 200 compositions during that time. It's an evolving/changing process. But, lately, instead of starting off by sequencing a short passage from classical music, as I did with my piece "Daydream" for flute and piano, where I used Debussy's "Reverie," and chose measures I did from AI sequences, lately, I've been working from original seeds generated by the sequencer itself. Here is a link to Daydream: th-cam.com/video/y53CjztrF0Q/w-d-xo.html The composition process relies on a total command of the use of MIDI software to construct the composition, change notes, manipulate phrases, and shape the performance quality as well as the composition. For example, sometimes AI will generate sequences where every note is the same volume. That needs to be changed. I made a video on how to use Allegro Music Transformer here: th-cam.com/video/6UqbNCabN3c/w-d-xo.html Just recently, I've started using the simplest tool to generate my 800 tokens long seed sequences. It can be found here: huggingface.co/spaces/asigalov61/Allegro-Music-Transformer You just select the length (in tokens) you want with the slider, and after previewing the seed, to see if it is something you want to fix up and use to sequence (doing "continuations"), it can be saved by clicking on the colored text denoting its size in the bottom right corner of the page under the MIDI pane. You can preview it in the WAV pane. After I've prepared the seed in MIDI software, because the quality of the seed determines the character and quality of the continuation sequences. Then I use the more complicated AI programs on GItHub to generated several, usually about 30, long continuation sequences. I listen to those, and choose phrases from them to built the piece. But I often change notes, even hundreds of notes sometimes. Lastly, here is a link to a composers video where he explains his process, when he was working in MuseNet some years ago. It is similar to mine: th-cam.com/video/ODasvwQFHKs/w-d-xo.htmlsi=R2N0PBnlh4KxXRDy

@@Timzart7 That is wonderful, thank you so much for taking the time to give me such a detailed reply!

Thanks Aldo!

I thought of a tip this morning. You as a person interested in AI music composition might find it useful. I'm reverting to Allegro Music Transformer lately to generate original seeds. And even though it uses the lighter-weight model, and is slower than GItHub, the Hugging Face version for seed generation works really well. The advantage is without any set-up required, or time restriction, it is easy to generate seeds on it while you're doing something else, and it can generate an 800 (size selected with slider) token seed in about 32 seconds. You can download a midi seed that you like with the colored text near the MIDI panel at the lower right/bottom of the MIDI panel. There is no sign-up or sign-in either, I think. huggingface.co/spaces/asigalov61/Allegro-Music-Transformer

With the release of Alex's latest, Quad Music Transformer, I started experimenting with a comparison in the last week. So, keeping in mind you listed them in the correct order of development, and speed, with each getting progressively faster, for my purposes, Giant is the one I am going to continue with after a few weeks of working with Quad. Quad is the fastest, but the sequences, especially the Improv (for original seed generation), contain too many chords, like choral music that have voices moving together. So, unless writing for chorus, Quad may not be the best choice. Quad might also be better for multi-instrument. I haven't experimented with that feature of it, which is different from the others. So, LAMC, Allegro, and Giant, are pretty indistinguishable, except Giant is fastest, by far. Especially when using the Composer Version, speed makes a huge difference. As far as being as good as MuseNet, I find myself spending a lot more time slicing and dicing, building a piece, than with MuseNet. Since I've posted over 200 AI pieces, so about 100 with MuseNet and a 100 with Alex's programs, maybe it would be easier for you to judge if the pieces are as good. Things like AI voice generation, including for singing, like ACE Studio, are advancing rapidly, but what I hate about some of those is they are expensive subscriptions, and they don't even let you just do $20 for a month. It usually has to be a one-year subscription. The one thing I'm sure about is how grateful I am that Alex even wrote all these things, because I could not tolerate AIVA and some others I've tried. AI is advancing so rapidly, if anyone hears of free software for music generation that produces output in MIDI, please call it to my attention, especially if it's on GitHub, Hugging Face, or free to use.

@@Timzart7 Do you have a tutorial on how to use the Giant Music Transformer (and the Composer version) or is it the same as LAMC?

@@aldothalerGiant is similar to LAMC and Allegro. Note that the procedure for loading an existing seed has changed, been simplified. It's all in-line with the other cells. You click on the arrow to load the custom seed, and a box pops below for you to locate that seed on your computer. So, you don't have to go through the directory on the left. And in the Composer Version, Giant lets you choose the length (in tokens) of the continuation you want to generate. For tighter control, generate a continuation sequence between 50 and 150 tokens, whereas for a longer continuation, you can go higher.

This is a bare windows 11 PC, do you know what i need to be able to run quad like was in Colab

No, I'm pretty sure you can't. That would be a good question for Tristan Behrens, who is a specialist in AI and has written some music programs for Github. On one hand, I know that Stanford, or some place, made versions of MLMs that can run on a standalone PC, and it seems like there could be one that generates music. I know that Google's AI Magenta music suite that runs on a PC does just very simple stuff compared to transformers like Quad and the others Alex has written. On Github, when your daily time runs out, you are given the option of connecting with your own GPU. I've done that. I have a very small GPU but it makes it really slow, if it runs at all. I need the Tesla T4 on their end for it to run.

MuseNet ran for 3 years, beginning in 2019. It worked on the GPT-2 system. Currently, it no longer works. There is no chance it will be released as a plug-in. I use the music transformer/sequencers on GitHub. (See the details under my more recent posts for links to those.) They are more complex to use than MuseNet. One day there may be sequencers that work as plug-ins, but I haven't found any yet.

I used a seed file that sounds like Chopin. The character of the seed determines how the continuation sequences sound, which are generated from it. You can get that seed by using a piece of a MIDI from a an existing Chopin piece you like. If you read the details of this, though, I had this old seed I had generated using the Chopin style -- you could choose styles -- of MuseNet. MuseNet allowed you to choose from a dozen styles and it could generate seeds in the style of that composer. Here is a piece I did, where the seed file was a handful of measures from Ravel's Alborada del Gracioso: th-cam.com/video/sUJxaqxOICo/w-d-xo.html

How long should the seed be? ​@@Timzart7

@@_otroll5393Between 200 tokens and 800 tokens, so, usually about 4 to 16 measures of music. You can use less or more. Sometimes I prepare two seeds from the same original seed, and experiment with continuing something from a different place in the music. There's the question of doing a continuation in the middle of a phrase, or at the end of a phrase. You can try both and see if one works better.

What seed did you use for this piece in particular?@@Timzart7

@@_otroll5393It was a seed generated by MuseNet with the "Chopin" style selected. MuseNet became non-operational in 2022.

Yes. You can compose three or four measures in MIDI for a seed, in the style of Chopin, or, just cut out between 15 seconds and 30 seconds of a Chopin MIDI piece you like, from the Chopin MIDI section of the Classical MIDI Archives (Free to join, but you get a limited number of downloads per day). So, prepare that in MIDI software. I've done several pieces that way, including multiples using the intro of Debussy's First Arabesque. I use the Original Version of Giant Music Transformer, but you can also try the Composer Version. Both start with loading a custom seed (after loading the setup cells), and doing "continuations." But how I do it in the Original Version is, I note how many tokens are in the original seed, let's say, 824, and in the first slider for the "continuations," I set that to 824 or a little over if there is no exact 824, and then that whole seed will be included in the continuation sequence, and it will sequence continuations. You'll want to do continuations of about 2500 tokens. You'll note that Giant Music Transformer allows you to do up to 16 continuation sequences, but doing continuations over 2000 tokens, I only do 3 at a time, because it can overload the RAM and you'll get the arrow on the screen turning to red if that happens. Lately, I use from 2 to 120 continuations to make a piece. Actually, the piece I'm working on now and should be out in a few days, Saxophone Elegy, is made from only 2 continuation sequences. I'm doing a lot of changes in. It's hard to assign a percent, because that varies from piece to piece. Anyway, there's another program called Giant Music Transformer Bulk Generator that overcomes the RAM overload problem. In that, instead of having to do continuations in a "batch" of 2 or 3, you can do 20 "cycles" of 2 batches, and get 40 continuations, without the RAM overloading. So, from these continuations, I audition them, and select a dozen or so to take out phrases. All the start out with the, let's say you used some measures of Chopin, and then I cut out the real Chopin for the final piece. You can use measures from continuations out of sequence. And yes, in MIDI software, I change phrases around, alter phrases, change notes, cut out extra beats. Sometimes I write my own measures in, or more often, alter pieces of phrases to, let's say, fashion an ending or intro. Finally, I adjust the performance quality, adding accents, tempo changes, better quality piano sound (you should always sequence using the base piano sound though, not a good piano sound), dynamics, and finally, sustain pedal for piano. The better your skill in MIDI, the more you can do with the raw sequences. Also, before sequencing something from the Classical Archives, I strip out the sustain pedal in MIDI software. That only takes about 15 seconds to do. That's why I add sustain pedal last to the final piece. When learning how to use this sequencing software though, start with smaller pieces (lower tokens sizes), just to experiment. It will be really fast that way, and you'll be able to tell what it is doing. If you have any questions, just ask. And tell me how it's going. If you like Chopin, one of my favorites of the 200+ AI pieces I've done, is the "Viral Diamond Etude." I call it that because it got 1000 views in one day after it was posted, one of only two pieces of mine that is done that. Usually, I'm getting only like 10 views a day on new pieces. This piece was done with the now dead MuseNet, by OpenAI: th-cam.com/video/SdjF6SPzxXQ/w-d-xo.html

It can output in MIDI or WAV, but yes, the input seed is MIDI only. It is using the TensorFlow and PyTorch libraries to sequence. I've never used the WAV output, since I've worked with MIDI for decades. The WAV is good for people that can't get a MIDI to play. I know if I just discovered this, I'd be off and doing it like crazy. But, I hope you get a chance to explore more of the 200 pieces I've done in the last few years, because there's a lot of variety.

Thanks, sir, I appreciate the reply. I have been following your tutorials and experimenting with things tonight. But beyond that, you have a talent for assembling and developing these ideas. Obviously some of that is an innate musical ability. I am trying to quantise some of the output as we speak. I haven't transcribed anything for 20 years and I really don't want to dust of the old skills if a computer can do it in 2 minutes! All the best

@@hywelmorgan3800 So, you work in music writing software, which is definitely one approach? And MuseScore is really good free music writing software for anyone out there who wants the best free one. When I started using AI, my initial impulse was to work in the musical notation view, but I found it impossible to quantize the MIDI files into proper musical notation. I mean it was like a mess, and would take me a month to straighten out the piece. One problem is I don't know any music writing software that well. I'd really like to just put all my pieces with the sheet music in the video, instead of selecting videos. Anyway, though, I work on the music in the track view, and the piano role view of MIDI software, alternating between those two modes, and don't use the notation display at all, either in the MIDI software or separate music writing software. After auditioning the raw MIDI files, to hear what ones I want to use, to select phrases from them, I plop all of those onto tracks in the MIDI software of a DAW. Just recently, I ran into a viewer who did a number of nice pieces using MuseNet and he had beautiful manuscripts for all his pieces. I wrote to him, asking what am I missing in this transcription/quantizing process of getting MIDI into notation. He said he paid people on Fivver to transcribe his MIDIs into notation, that it took him a really long time to do it, too. But, maybe if I spent as much time using MuseScore, I'd be better at it. In the distant past, I used Finale and Sibelius for a little composing, but never got a real mastery of using music writing software. So, my experiences in trying to transcribe things were frustrating. Anyway, if that goes well for you, working in musical notation, and you have any tips on quantization, I'd be interested in what you learn in going through that process. It was not especially intuitive for me to use the piano roll view so much to compose, but now that I'm used to it, it allows completely control over every note and I'm okay with it. And when AI produces so many polyrhythms, I don't have to strain my brain trying to figure out the note values and time signatures. That way, I can just concentrate on how it sounds.

@@Timzart7 Yes, I'm messing around in MuseScore. Messing is probably apt. I suspect the Fiverr route is probably easiest depending on the cost. Just listening to some seeds from Giant but 99% of them are garbage. I guess that's all part of the game. I've subscribed anyhow and will follow your progress with eagerness. All the best!

@@hywelmorgan3800 Don't be afraid to ask a question if something pops up while you're working and you want to know something specific about a setting in the software, or what a term means. Also, it's fun do do continuations using a seed you write yourself, which can be as short as a measure or two, or by selecting a seed from piece where you like the style, and that seed can be anything from 10 seconds of music to up to about 45 seconds.

@@Timzart7 Thanks, Timzart. I look forward to sharing a few messages and ideas with you in the future! All the best!

Los Angeles Music Composer, Allegro Music Transformer, and Giant Music Transformer are all similar, by the same author, so if you use one -- I started out with Los Angeles Music Composer myself -- you can use the other, with Giant being the most recent and the fastest to generate sequences.

I'm glad you've commented! I started listening to your fine pieces and will listen to all of them. I've done searches, wanting to hear what others have done with MuseNet, and have never encountered someone like myself, who takes it seriously, except Alex, who is a genius and wrote all the software I'm currently using. He started with MuseNet, too. His MuseNet pieces were all in a file on Reddit. Of course, I've found people on YT who posted their MuseNet compositions, but few of them have done a lot to refine the performance quality of the pieces, and on some pieces, I auditioned as many as 50 sequences for the "next measure" although a good enough one was usually in the 4 it generated at a time. I liked MuseNet because I would work in 3 to 4 tabs at once, and for my piece "Mirage" (on YT, search Timzart7 Mirage) that was probably done from 6 sequences, all from the same intro, which was an original generated by MuseNet. I tried working in 10 tabs at once. Three was enough though, because no waiting. MuseNet went down for good on Dec. 12, 2022, after more than three years. It was terribly underutilized, and got a review from some professor at MIT that made me want to scream. It was really amazing software, the way note length was built into pieces, so I didn't have to go through and put in sustain pedal, like I do now. I still miss it. Lately, I use Alex's program Giant Music Transformer Bulk Generator, to take a seed/intro I've prepared, usually something that was generated by AI in Giant, and generate 60 long continuations of 3000 or so tokens length, and listen to them all, and select a dozen from which to pick phrases. I also sometimes use another tool called "inpainting." In MuseNet, I rarely used mostly 50 to 150 for the sequence length, for the continuation. I didn't discover any AI music software until 2022, when I began using MuseNet in the last four months of its existence. I was following various LLM blogs and didn't realize AI was used in music, for free anyway. Starting with a some programming computers in 1971 when I was a 17 year old music student in college, I never became a good programmer, but tried, taking courses in several languages in the decades following. I'm a coding dunce. Then I went to Europe to study music and ended up leaving and pursuing a career unrelated to music or computers for some years. But I came back to music. I always liked music composition in school, but just have dabbled with composition throughout my life. But like you, being able to compose makes working with AI easier, since like you said, at present, it has no sense of the whole. It's other characteristics are it puts in a lot of extra beats, and phrases tend to be too long, with very few rests. It usually can't write an ending. I do those. The problem for more causal users is that MIDI software or music writing software takes months to learn. I was never that good with MIDI, and it has only been since I started using MuseNet that I've learned how to control everything. I'm still terrible at music writing software, like MuseScore, and am going to ask you a question on how the you produce such great manuscripts, like from your MuseNet MIDI pieces, how long that takes. I'm doing something wrong. But my main problem is I'm terrible at music writing software, very slow, and I'm sure you're totally proficient in it. At present, I could take a piece even as simple as this, and it would probably take me weeks to straighten out all the notes to produce a manuscript. I quantize and understand the principles, but these pieces look like a nightmare mess in notation. About the familiarity of this tune, I thought that, too, but couldn't put my finger on what it was. To answer your question, I think it would be rare for the AI to pull out a full melody from a training library. In my 200+ pieces, with only a few exceptions, I usually remove any intros I used from Debussy or whatever in the final piece. And I remember one Chopin style piece I started with a few measures from an Etude, and there was a passage that had the same arpeggio as in the real etude, and I just don't use those. I've used portions of Debussy's First Arabesque for several pieces. And I've used parts of Clair de Lune, where the sequences passage is so different from the original, I can hardly tell what part of the the original is being used myself. Lately, though, I use original seeds/intros. I've also written a few of my own intros. I did a version of Fur Elise also. I like yours a lot! But of my MuseNet pieces, I think Mirage and Viral Diamond Etude (which is the second version of the version that went mini-viral, with over 1000 hits in one day) are pretty unusual. The Diamond Etude was an early MuseNet piece and it was robotic and all the notes same volume, so I went to great effort to make it sound more natural. Most of my latest posted pieces get like 10 views a week now, so I'm hardly the viral AI sensation. I'm just doing this for fun in my retirement and to maybe help others who want to play with AI, if they like what they hear. Lastly, for anyone reading War & Peace here, who wants to try the currently available AI sequencing software, I've made tutorials of Allegro Music Transformer and Los Angeles Music Composer.

Do you want original seeds or continuations? In the description under the video, I give a link to my tutorial on Allegro Music Transformer, 'Original Version" which is similar to the Original Version of Giant Music Transformer. However, the bulk generator has a few small differences. Instead of saving files from the file directory, there is another cell to generate after you've, for example, performed a continuation on a seed (which is identical to the way a continuation is done in Allegro Music Transformer, Original Version). After the cell is done generating, you click on the cell below and it saves all the files and zips them and downloads them, or, provides a name you want to save it under. I believe you wrote once that your computer has only 8GB of RAM. This is where the Bulk Generator could be very useful for you. In the "continuation" cell, you could generate 20 generation cycles of a batch of two 2500 tokens long continuation sequences, and end up with 40 long continuations without overloading your RAM. Again, though, to save those continuations, you have to generate the cell after it which will zip and save the 40 continuation sequences to your download directory on your computer. I may do a video tutorial on the Giant Bulk Generator. I'm just not in the mood for it right now. So, the whole purpose of the Bulk Generator is so you can generate seeds (using Improv) or continuations in bulk, instead of being limited to 16 at a time. Even though I have 16GB of RAM, the regular Giant Music Transformer could give me the red button overload, if I generated a batch of 6 or more continuations that were really long. This morning, for making a piece, I generated 60 continuations of 2800 tokens length, which took one hour to do. Still, it gives me a lot of versions to choose from for making a piece.

@@Timzart7 continuations

So, you have a MIDI clip that is multi-track, each track assigned to a MIDI channel/instrument, and you want to upload it to the software and have AI continue the piece on those tracks? The following free software (Python in a notebook) can do that: github.com/asigalov61/Pentagram-Music-Transformer I have not done a tutorial for Pentagram, but have done one for the "Original Version" of Allegro Music Transformer which you can watch here: th-cam.com/video/6UqbNCabN3c/w-d-xo.htmlsi=qiAX4Xp-ZaD3yPQ1 So, do the tutorial for Allegro, paying attention to the "continuation" feature, but then try working in the "Original Version" of Pentagram, which puts the instruments on separate tracks. I've used it only once. For the first time learning AI sequencing software for that works with MIDI like this, especially in MI (multi-instrumental), start with a simple two-track, two instrument piece, rather than trying to feed in 20 tracks of something 5 minutes long, like an orchestral arrangement. A simple piece will process faster, and you'll be able to learn what the various parameters do and how to use the software. Keep in mind, AI is at a collaborative phase right now. It's a tool. It really helps, at this point, to have strong knowledge of MIDI software so you can clean up things in post-process. Most people, unless you are a very advanced MIDI user and understand everything about channel assignment and the programmer-level of how MIDI works, usually learn AI MIDI software using just piano, or even a single instrument. Then branch out to multi-instrument. I've posted over 200 AI MIDI compositions in the last year and a half, and less than 10% are multi-instrumental. And when I did those, I created the tracks myself and separated the lines manually in MIDI software. I wasn't using the multi-instrumental capabilities of a program like Pentagram. These AI MIDI programs that work off of training libraries like PyTorch and TensorFlow can be glitchy, and produce a lot of garbage. You have to be fascinated by AI MIDI sequences and determined to learn the software. Once you learn the software, it can be a great tool for a song writer or composer, who gets stuck, and just wants ideas to continue a certain phrase, or for whole compositions, as I use it. One example of a glitch is you can select "piano only" and some of the continuation sequences will generate in multiple random instruments. This happens maybe one out of 20 times.