Generative Ai (ChatGPT)

[DE.]

There's a lot of open source stuff around at the moment, but it's still a bit of a wild west in terms of commercial tools. At this point for every commercial tool that's available, there are open source alternatives which do the same job, even if it's slightly more difficult at first to get to grips with. 

I'm not sure there's a huge amount of useful commercial stuff out there right now (that doesn't have a similar quality open source alternative), but for open source AI products you basically just need a decent understanding of GIT/GIThub, a basic knowledge of how to install and put in commands for Python, and the concentration levels to watch 10-20 min YouTube tutorials explaining how to get going. 

I've got some incredibly impressive results using some open source art (Stable Diffusion) and music/voice cloning (Ultimate Vocal Remover/RVC-Beta) AI, and it didn't take very long to get going with any of them. Stable Diffusion needs some pretty big files to be downloaded though. None of these AI tools are likely to be useful for what I'm doing now or in the near future, but I think getting used to how tools like this operate is still worthwhile. 

[RoverDom]

https://www.ben-evans.com/benedictevans/2023/7/2/working-with-ai?utm_source=tldrai

 

Fantastic read and perfectly articulates what AI will do and its impact on jobs etc

[RoverDom]

After recommendations for AI photo editing. Specifically I upload a photo I've taken and ask it to amend the photo in some way e.g. remove the fruit bowl from the table- that kind of thing. @DE.do you know of anything? 

Edited March 9 by RoverDom

[DE.]

What is your PC/laptop setup like? If you have a decent setup with a relatively good graphics card you can use a local install of Stable Diffusion to edit pictures. There's a step-by-step guide below which advises how to install the AUTOMATIC1111 extension, which is what I've used in the past:

https://stable-diffusion-art.com/install-windows/

Once you've installed it, you can use the img2img tab to insert your original picture and then use the prompts to change it, or inpainting to remove certain items. It can take a little time to fiddle with settings and prompts to get what you want, but generally it does a good job. 

There are other options such as Google Collab if you don't have the local power to run an AI generative tool - the program itself is installed on a remote server, so you're only using your browser. Similar to cloud gaming.

https://www.unlimiteddreamco.xyz/articles/how-to-run-stable-diffusion-in-google-colab/

Otherwise there are the more traditional methods of using Photoshop or GIMP to edit photos. If it's something relatively simple like removing items from a table, then using something like the clone tool to edit the table surface over the item on the table can sometimes give perfectly good results. Depends how complex it needs to be.  

 

[Mike E]

I think the Glasgow Willy Wonka Experience shows that AI has a way to go before it can overtake actual creativity.

[RoverDom]

11 hours ago, Mike E said:

I think the Glasgow Willy Wonka Experience shows that AI has a way to go before it can overtake actual creativity.

As long as its good enough to rip off my cowboy of an estate agents photos, it'll do for now 🤣

[DeeCee]

I've got an AI camera on my new phone - wtf?

 

[bluebruce]

On 27/06/2023 at 17:35, DE. said:

 

I've got some incredibly impressive results using some open source art (Stable Diffusion) and music/voice cloning (Ultimate Vocal Remover/RVC-Beta) AI

This is one of the things that worries me most personally...I'm an actor and I decided that this year I'm going to start doing a bit of voiceover work. Not as a main career unless it took off more than I expect, but still. I'm concerned that (as is typical for my luck) I've chosen pretty much the exact worst time to do this. I'm going to have to be careful reading contracts and making sure I can't be cloned (though it will probably happen anyway), but moreover this seems to be an existential threat to the voiceover industry. You already hear a lot of clearly AI-generated voices on youtube videos, and maybe some that are less obvious. Of course, for a time the work at the top end will still be mostly real people as there's a certain level of artistry that an AI can't match yet, but I think the bottom end I'm starting out at will be swamped with cheap AI voices fairly quickly. If your budget is small why would you use an actor when you can just get an AI to generate it? It won't be as good but it will probably be close enough to suit their purposes.

Give it, I dunno how many but let's say 30 years or probably a lot less, and even screen actors will be replaceable with AI generated images and sound without anyone being able to tell the difference. The question then will become whether people are happy to watch that sort of content or will insist on seeing real people. Will they even know which they're watching? I guess with celebrities maybe, but how will you know that actor didn't just phone it in and give the company permission to use them as an AI instead and bank the money? And how will actors reach celebrity status in the first place if they are competing with much cheaper AI that do whatever they're told at all times?

I imagine it will follow the usual pattern of technological change- at first lots of people, especially older generations, won't want any truck with watching films like that, then over time it will just become the norm.

Most of my work is live, and that's not going anywhere any time soon (you'd need expensive hyper realistic androids or holograms that haven't even been made yet), so I'm not overly concerned, but it does make you wonder. Tech is wonderful and I generally embrace it, but art doesn't really feel like an area we should give over to machines. That's before we get into the whole Skynet kind of thing...

[DE.]

It's genuinely scary how accurate voice cloning has become, so I can understand the pessimism. What is noticeable is that AI struggles with natural accents and general dialects. Like, 95% will be fine but then there's 5% that just sounds wrong. Not that the voice itself is wrong, but the inflection doesn't quite match what you'd expect to hear.

From what I've read this is (at the moment) being solved by having somebody provide the vocal rhythm and cadance, and then putting the AI over the top of it to change the sound to the voice of the person you're cloning. So you don't have to actually sound like the person in question, you just have to be able to match their accent and general delivery of sentences. So, in theory a voice over studio could license out a famous person's voice and then hire a random voice actor to deliver the lines with the AI voice being laid over the top. Ultimately I imagine the technology will advance to the point where the VA in the middle is no longer required, but I can see that being the interim process. 

To be honest in terms of AI I mainly just use ChatGPT and Ultimate Vocal Remover these days - the former to help with coding and the latter for musical endeavours. Both very cool and in terms of the latter, something I wish had existed 14 years ago when I could really have made use of it. But alas. Still very cool to be able to create instrumental versions of basically any song I want within 2 or 3 minutes, as well as being able to play around with the vocal stem. 

I don't really use Stable Diffusion for anything these days (although I'd probably use it if I wanted to create temporary front/back pages for any books I'd written that I wanted to potentially do something with - whereas before I'd have been asking someone on fiverr or something) and most of the other AI related software is a bit too intensive for my current setup. Laptop has an RTX 2070 which is reasonable enough, but the heat generated as programs like RVC-Beta run is too much of a concern for me. Not so bad on the tower, but that's only got a GTX 1070 - still VR capable which is nice, but in terms of AI somewhat slow.   

Edited April 23 by DE.

[bluebruce]

1 hour ago, DE. said:

To be honest in terms of AI I mainly just use ChatGPT and Ultimate Vocal Remover these days - the former to help with coding and the latter for musical endeavours. Both very cool and in terms of the latter, something I wish had existed 14 years ago when I could really have made use of it. But alas. Still very cool to be able to create instrumental versions of basically any song I want within 2 or 3 minutes, as well as being able to play around with the vocal stem.

That's actually something I'm interested in myself, as I've wanted instrumentals for a relatively underground artist for something I intend to do later in the year and I don't think I'll be able to find them ready-made. I've been assuming something like that must exist, but hadn't got to the point of looking it up, so appreciate you mentioning one you've worked with that can be relied on. I see it's free too! Can it be used to keep some vocal elements in place, like if you wanted to keep the chorus, or some kind of 'background' voices etc?

[DE.]

34 minutes ago, bluebruce said:

That's actually something I'm interested in myself, as I've wanted instrumentals for a relatively underground artist for something I intend to do later in the year and I don't think I'll be able to find them ready-made. I've been assuming something like that must exist, but hadn't got to the point of looking it up, so appreciate you mentioning one you've worked with that can be relied on. I see it's free too! Can it be used to keep some vocal elements in place, like if you wanted to keep the chorus, or some kind of 'background' voices etc?

Yeah absolutely, you just need an audio editing tool. I tend to use something like Cool Edit 2.1 which is ridiculously old these days, but if you're literally just wanting to splice two audio tracks together it's perfectly capable. Programs like Adobe Audition are obviously better, but it depends how far you want to go with it really.

Anyway, it's pretty simple. Download the program, load the track you want into the input, select your output, then tweak the settings to your liking. I've got mine set up as below:

But YMMV depending on the type of music you're working with. I primarily extract rock/metal and this works pretty well, but not sure how it would fare with other genres. 

Anyway, it'll save a vocal and an instrumental stem. If using the above setup make sure to check the HP-UVR extracts that are created - I often find these sound better than the fully ensembled output. 

Once you have your instrumental and vocal files output and you're happy with them, you're sorted. Open up your audio editing program, insert the instrumental to channel one and the vocals to channel two... and then do whatever you please. Audio volume between both tracks should be perfect so no need to mess around with those, unless you want to lower or raise volume. 

[DE.]

So, Google Notebook LM now has the ability for you to upload a PDF ot TXT file and have AI generate a 6-10 min conversation in a podcast style discussing the contents. I'm not floored by too much in the AI sphere anymore, but this is absolutely amazing. It sounds totally natural and real.

https://notebooklm.google.com/

Really scary - and this is just publicly released stuff. One can only imagine what exists behind closed doors at google, OpenAI, etc.

[RoverDom]

On 18/09/2024 at 20:40, DE. said:

So, Google Notebook LM now has the ability for you to upload a PDF ot TXT file and have AI generate a 6-10 min conversation in a podcast style discussing the contents. I'm not floored by too much in the AI sphere anymore, but this is absolutely amazing. It sounds totally natural and real.

https://notebooklm.google.com/

Really scary - and this is just publicly released stuff. One can only imagine what exists behind closed doors at google, OpenAI, etc.

I think what scares me a little bit is not what AI can do currently (although some is mind blowing) but that we're really only at the beginning. What does it look like in 10 years time? 

[StubbsUK]

I don't find it scary, I find it disappointing. AI should be doing all the crap jobs and leaving us humans to do the fun stuff like art, music and writing, but instead it largely seems to be the other way round.

Of course, the issue is the art, music & prose being generated is just that, it's using stuff humans have already done and recombining it in a way that looks different enough to be passable as something new. There's nothing intelligent about the current generative models ... yet.

[DE.]

You can generally always tell purely AI generated prose. It has a very specific kind of structure to it that's recognisable immediately if you've used AI enough in the past. Just that uncanny valley sensation of almost being real, but not quite.

What I have enjoyed doing recently is letting AI generate insights into things that I've written. It digs out themes and observations that I hadn't even considered, and helped me refine my focus when adding to that narrative in future. For the analysis side of things, AI is very good. Also for repairing or reconstructing things. From a generative perspective, mostly derivative at this point in time. Fun, but kind of soulless and loses its appeal quite quickly. 

[Norbert Rassragr]

Hopefully Skynet will be too distracted by cat videos to take over the world.

[DE.]

For anyone interested in generative AI for music, Suno AI has taken pretty huge steps since I last used it a few months back. If you download it on your phone it lets you use the new v4 model for a while, although eventually you have to revert to the 3.5 model (both are pretty good though). I'm honestly amazed at how far the technology has come since I last used it. Granted it's better at some genres than others. It seems to do a really good job with rap/hip hop. Pretty good with pop, although it's obvious the training data has a lot of autotuned vocals and that can be annoying. I like the rock/metal songs it creates, but the production and mixing is a bit off. 

An example of a couple of songs I created below - I provided the lyrics and general song direction (verses, bridges, solos, etc) whilst the AI produced the music, vocals and vocal melodies. 

 

On the first one the autotune issue is present at times, whilst on the second one the production/mixing issues are evident. Nonetheless, I imagine someone with a lot more time and software could improve both significantly. I imagine at this early stage most of what it produces is ultimately still pretty formulaic and much like ChatGPT would be increasingly easy to spot if mass produced, I haven't really listened to enough AI generated music to know for sure.

I sent some of the stuff I'd made to a friend in the music business, and he was both impressed and horrified by what was possible now with AI. Obviously from the perspective of me - as an individual creating music for personal enjoyment that I'd never pay anybody to create - it doesn't harm anyone, but you could see this kind of AI being used in ways which really gut the industry of various professionals involved in the process. It's especially scary when you consider we're still in the embryonic stages of generative AI as a whole. 

Edited December 13 by DE.