Loopy Pro: Create music, your way.

What is Loopy Pro? — Loopy Pro is a powerful, flexible, and intuitive live looper, sampler, clip launcher and DAW for iPhone and iPad. At its core, it allows you to record and layer sounds in real-time to create complex musical arrangements. But it doesn’t stop there—Loopy Pro offers advanced tools to customize your workflow, build dynamic performance setups, and create a seamless connection between instruments, effects, and external gear.

Use it for live looping, sequencing, arranging, mixing, and much more. Whether you're a live performer, a producer, or just experimenting with sound, Loopy Pro helps you take control of your creative process.

Download on the App Store

Loopy Pro is your all-in-one musical toolkit. Try it for free today.

"A.I." (Machine Learning Algorithms) To Generate Art

1…1719212223

Comments

  • supadomsupadom
    edited February 2023

    @echoopera said:

    @supadom said:
    I don’t know why but this whole AI generated art reminds me a lot of my neighbour’s astroturf.

    Imho it’s no different than Sampling in Music. It all depends on how you use the output it gives you. If you just rely on the initial image then yes, it is like astroturf…but if you slice, dice, resample and build upon parts and pieces of the output you create an entirely new landscape to enjoy. I’m taking this opportunity to get into my Robert Rauschenberg mode with all of these tools.

    That’s my view of it after thinking about it over the last few months. The long view is that for visual artists who want to use the system to iterate on their preferred styles and motifs it is an amazing tool and this is the lens i am seeing it through. I see it in the same light as i see Riffer, Fugue Machine, PlayBeat3, Scaler 2—a tool to push me in to new places based on my intention.

    They need to work out the legality of it all though…but in the meantime, it’s a vast Sample Crate i am enjoying to create visuals with when inspiration strikes. I’m using it to generate a ton of textures and elements i can use for years. 😉

    I guess my reaction is comparable to the critics of Warhol when he put pop art out there. I like the stuff that you’ve posted but can’t help asking where is the line between machine and the human.
    Traditionally fine art was about the manual skill, capturing the beauty, now, god knows, it’s become so hazy that I don’t really want to get into it.
    I guess it’s for the maker of the output to decide how honest they feel about what they are putting out.
    How does one feel having written a song initiated by a press of a button and then polishing it off as opposed to being written based on the feeling generated by the heart/brain collaboration.
    Yes one can concoct a heart/brain/ai axis and make it ‘fly’ but I, personally, just don’t buy it.

  • supadomsupadom

    @echoopera You make comparison to sampling. Well, it’s a spectrum. There’s those who mush 2 songs together DJ style and call themselves producers. There’s other who use the sampling form in a more subtle way.
    Who is the judge here?

  • echooperaechoopera

    @supadom said:
    @echoopera You make comparison to sampling. Well, it’s a spectrum. There’s those who mush 2 songs together DJ style and call themselves producers. There’s other who use the sampling form in a more subtle way.
    Who is the judge here?

    I’m equating it to how i use it. I can only judge myself and talk about my process. I have no control over what others do or think on the matter.

    But i do have my opinions of course 🤪

  • knewspeakknewspeak

    The Manchine is becoming. Like the tree will it’s fruit be poisonous?

  • echooperaechoopera
    edited February 2023

    The machine has become more than a tree. It is a vast forest and network with many offerings.

    Some seek the forbidden while others place composts on the heap to yield new varietals and opportunities.

    If life and history has taught us anything, what Prometheus brought to humanity changed the world. How we use the flame is another matter altogether.

    “Reap what you sow” is an essence of our nature so we should bear this in mind with all things we touch in this life.

  • knewspeakknewspeak

    @echoopera said:
    The machine has become more than a tree. It is a vast forest and network with many offerings.

    Some seek the forbidden while others place composts on the heap to yield new varietals and opportunities.

    If life and history has taught us anything, what Prometheus brought to humanity changed the world. How we use the flame is another matter altogether.

    Reap what you sow is the essence of our nature so bear this in mind with all things we touch.

    What we sow now will bear fruit for our children, to those, yet without existence, all of them, we have a duty of care.

    Prometheus brought knowledge, both good and bad, it’s implementation without wisdom could well be the ultimate stupidity.

  • CarnbotCarnbot

    @AudioGus said:

    @Carnbot said:

    @AudioGus said:

    @Danny_Mammy said:
    Just like audio sampling one day they will be able to trace the original content the AI scanned and copyright claim the F%^k out of everyone. Be prepared.

    Diffusion rendering is nothing like audio sampling. Not to say that it has totally ethical roots of course, and while there are a few rare cases of overfitting in even the best ML models (where you can tell the source material) for the most part in 99.99999% of diffusion renderings things are completely, effectively laundered into oblivion.

    That's not strictly true, the data is in the model and can be replicated and reconstructed quite accurately but it's never asked to usually or it happens accidentally, since you have very little control over the data with text to image/audio apps. So while it's not exactly like traditional sampling it's a different type of sampling, eg data sampling, but it's closer to it than not.

    The images cannot be replicated and reconstructed unless they are overfitted, for example with many extraneous duplicates present in the original dataset when the model was trained. In the case of Stable Diffusion it was trained on billions of images and the model is only 4gigs. There is no way you can replicate and reconstruct those billions of images from a 4 gig file. There are certainly some examples of overfitting there but the vast overwhelming majority are vapor.

    The data is compressed with other data where there are similarities, so it won't come out exactly the same, but quite close. that's what it's doing really, removing the data it doesn't need from the masses of images to replicate a similar pattern and reduce the model, so it's creating relationships between all the data it's analysed. I think that looking at it like a data visualisation explorer is a good perspective as shown in the video above.

  • AudioGusAudioGus
    edited February 2023

    @Carnbot said:

    @AudioGus said:

    @Carnbot said:

    @AudioGus said:

    @Danny_Mammy said:
    Just like audio sampling one day they will be able to trace the original content the AI scanned and copyright claim the F%^k out of everyone. Be prepared.

    Diffusion rendering is nothing like audio sampling. Not to say that it has totally ethical roots of course, and while there are a few rare cases of overfitting in even the best ML models (where you can tell the source material) for the most part in 99.99999% of diffusion renderings things are completely, effectively laundered into oblivion.

    That's not strictly true, the data is in the model and can be replicated and reconstructed quite accurately but it's never asked to usually or it happens accidentally, since you have very little control over the data with text to image/audio apps. So while it's not exactly like traditional sampling it's a different type of sampling, eg data sampling, but it's closer to it than not.

    The images cannot be replicated and reconstructed unless they are overfitted, for example with many extraneous duplicates present in the original dataset when the model was trained. In the case of Stable Diffusion it was trained on billions of images and the model is only 4gigs. There is no way you can replicate and reconstruct those billions of images from a 4 gig file. There are certainly some examples of overfitting there but the vast overwhelming majority are vapor.

    The data is compressed with other data where there are similarities, so it won't come out exactly the same, but quite close. that's what it's doing really, removing the data it doesn't need from the masses of images to replicate a similar pattern and reduce the model, so it's creating relationships between all the data it's analysed. I think that looking at it like a data visualisation explorer is a good perspective as shown in the video above.

    I have seen some bad videos that make simplistic descriptions of the diffusion rendering method that imply that you can completely reverse render every dataset from any model. With a small experimental dataset/model sure why not, but with a practical one with a massive properly curated and trained dataset, it is impossible. Again, only things that are overtrained by having too many duplicates in the source dataset will exhibit this. Models that can do this are essentially flawed.

    When you finetune SD using dreambooth with your own single image you need to crank up the training steps so much that it bleeds into other concepts in order to be legible, at which point you have effectively broken the model. To say that every image that it is trained on can be reproduced is simply not true, unless you make a bad, impractical and unusable model to specifically support this premise.

  • echooperaechoopera
    edited February 2023

    @knewspeak said:

    @echoopera said:
    The machine has become more than a tree. It is a vast forest and network with many offerings.

    Some seek the forbidden while others place composts on the heap to yield new varietals and opportunities.

    If life and history has taught us anything, what Prometheus brought to humanity changed the world. How we use the flame is another matter altogether.

    Reap what you sow is the essence of our nature so bear this in mind with all things we touch.

    What we sow now will bear fruit for our children, to those, yet without existence, all of them, we have a duty of care.

    Prometheus brought knowledge, both good and bad, it’s implementation without wisdom could well be the ultimate stupidity.

    Now we’re just talking in circles. 🤪 Waxing poetic i think they call it.

    What are you saying?

    Unplug the system?
    Regulate it?
    Take it to court and resolve the IP issue?
    Treat it like a streaming service and pay all the artists referenced in the image making outputs when their signal is used?
    All of the above?
    😉

  • supadomsupadom
    edited February 2023

    @echoopera said:

    @supadom said:
    @echoopera You make comparison to sampling. Well, it’s a spectrum. There’s those who mush 2 songs together DJ style and call themselves producers. There’s other who use the sampling form in a more subtle way.
    Who is the judge here?

    I’m equating it to how i use it. I can only judge myself and talk about my process. I have no control over what others do or think on the matter.

    But i do have my opinions of course 🤪

    Absolutely, I’m not pointing fingers at you at all, or at nobody in particular, in fact. I’m just musing about dilution of skill, of which (skill) there’s plenty of around, no doubt. Although I find it rather like the fake news situation where it’s hard to tell one thing from another.

    I’m not saying that IA cannot be used positively, just it confuses the shit out of me as it seems to be breaking some well established hierarchies. Maybe that’s what progress is supposed to do, I don’t know…and maybe that’s how people’s art takes the revenge on the privileged. As I said, I’m well confused by it. Especially when my friends some of whom are decent songwriters tell me how they’re using IA in times of creative drought.

    In school, the everybody gets a medal always rubbed me the wrong way but it sure has its merits.

  • NeuMNeuM

    @supadom said:

    @echoopera said:

    @supadom said:
    @echoopera You make comparison to sampling. Well, it’s a spectrum. There’s those who mush 2 songs together DJ style and call themselves producers. There’s other who use the sampling form in a more subtle way.
    Who is the judge here?

    I’m equating it to how i use it. I can only judge myself and talk about my process. I have no control over what others do or think on the matter.

    But i do have my opinions of course 🤪

    Absolutely, I’m not pointing fingers at you at all, or at nobody in particular, in fact. I’m just musing about dilution of skill, of which (skill) there’s plenty of around, no doubt. Although I find it rather like the fake news situation where it’s hard to tell one thing from another.

    I’m not saying that IA cannot be used positively, just it confuses the shit out of me as it seems to be breaking some well established hierarchies. Maybe that’s what progress is supposed to do, I don’t know…and maybe that’s how people’s art takes the revenge on the privileged. As I said, I’m well confused by it. Especially when my friends some of whom are decent songwriters tell me how they’re using IA in times of creative drought.

    In school, the everybody gets a medal always rubbed me the wrong way but it sure has its merits.

    Frankly, it doesn't surprise me that your songwriter friends should use these tools. They're tools. Artists and musicians are typically the first to adopt new things, and I see no difference with machine learning tools.

  • CarnbotCarnbot

    @AudioGus said:

    @Carnbot said:

    @AudioGus said:

    @Carnbot said:

    @AudioGus said:

    @Danny_Mammy said:
    Just like audio sampling one day they will be able to trace the original content the AI scanned and copyright claim the F%^k out of everyone. Be prepared.

    Diffusion rendering is nothing like audio sampling. Not to say that it has totally ethical roots of course, and while there are a few rare cases of overfitting in even the best ML models (where you can tell the source material) for the most part in 99.99999% of diffusion renderings things are completely, effectively laundered into oblivion.

    That's not strictly true, the data is in the model and can be replicated and reconstructed quite accurately but it's never asked to usually or it happens accidentally, since you have very little control over the data with text to image/audio apps. So while it's not exactly like traditional sampling it's a different type of sampling, eg data sampling, but it's closer to it than not.

    The images cannot be replicated and reconstructed unless they are overfitted, for example with many extraneous duplicates present in the original dataset when the model was trained. In the case of Stable Diffusion it was trained on billions of images and the model is only 4gigs. There is no way you can replicate and reconstruct those billions of images from a 4 gig file. There are certainly some examples of overfitting there but the vast overwhelming majority are vapor.

    The data is compressed with other data where there are similarities, so it won't come out exactly the same, but quite close. that's what it's doing really, removing the data it doesn't need from the masses of images to replicate a similar pattern and reduce the model, so it's creating relationships between all the data it's analysed. I think that looking at it like a data visualisation explorer is a good perspective as shown in the video above.

    I have seen some bad videos that make simplistic descriptions of the diffusion rendering method that imply that you can completely reverse render every dataset from any model. With a small experimental dataset/model sure why not, but with a practical one with a massive properly curated and trained dataset, it is impossible. Again, only things that are overtrained by having too many duplicates in the source dataset will exhibit this. Models that can do this are essentially flawed.

    When you finetune SD using dreambooth with your own single image you need to crank up the training steps so much that it bleeds into other concepts in order to be legible, at which point you have effectively broken the model. To say that every image that it is trained on can be reproduced is simply not true, unless you make a bad, impractical and unusable model to specifically support this premise.

    It will depend on how unique a piece of input data is (most are not that different from another one if the dataset is huge) if it can be recreated accurately or not. It's amazing how close they can get to the originals if the correct coordinates in the model map is found, as close as a compressed jpeg is to an uncompressed image, but it will depend on the kind of image or data being replicated.
    Is it similar to sampling? Yeah I'd call it mega-sampling, more sampling than a human can ever do in a short space of time. It's purpose in mega-sampling is to find the correlations and biases in the dataset. In the human created image datasets it reveals the human image biases like a map. It's very easy to see that this mega-sampling is similar to normal sampling but on an industrial massive scale if the model is big, it's an infographic of the human input data. So unless using all data which you've input yourself, when "generating" material with these models it's much closer to curating than creating.
    If I try to create something in my mind using these models it never comes out how I want, it's always biased towards the models patterns, because I'm exploring a dataset. If it comes out how you want, then it's confirming your biases align with the model.

  • NeuMNeuM
    edited February 2023

    As far as I’m concerned, these new tools represent an opportunity for established artists. I know there are some lawsuits going forward by artists suing to prevent their work being “scraped” to train these models, however I see this as possibly the best time for artists to get ahead of this and LICENSE their work to these art generators to better train them so they are more effective. The best artists will continue to create with or without these new tools, but they could be earning additional revenue at the same time with smart licensing arrangements. In countries where they do not respect licensing and intellectual property rights they’re going to steal and use whatever they want anyway.

  • AudioGusAudioGus
    edited February 2023

    @Carnbot said:

    @AudioGus said:

    @Carnbot said:

    @AudioGus said:

    @Carnbot said:

    @AudioGus said:

    @Danny_Mammy said:
    Just like audio sampling one day they will be able to trace the original content the AI scanned and copyright claim the F%^k out of everyone. Be prepared.

    Diffusion rendering is nothing like audio sampling. Not to say that it has totally ethical roots of course, and while there are a few rare cases of overfitting in even the best ML models (where you can tell the source material) for the most part in 99.99999% of diffusion renderings things are completely, effectively laundered into oblivion.

    That's not strictly true, the data is in the model and can be replicated and reconstructed quite accurately but it's never asked to usually or it happens accidentally, since you have very little control over the data with text to image/audio apps. So while it's not exactly like traditional sampling it's a different type of sampling, eg data sampling, but it's closer to it than not.

    The images cannot be replicated and reconstructed unless they are overfitted, for example with many extraneous duplicates present in the original dataset when the model was trained. In the case of Stable Diffusion it was trained on billions of images and the model is only 4gigs. There is no way you can replicate and reconstruct those billions of images from a 4 gig file. There are certainly some examples of overfitting there but the vast overwhelming majority are vapor.

    The data is compressed with other data where there are similarities, so it won't come out exactly the same, but quite close. that's what it's doing really, removing the data it doesn't need from the masses of images to replicate a similar pattern and reduce the model, so it's creating relationships between all the data it's analysed. I think that looking at it like a data visualisation explorer is a good perspective as shown in the video above.

    I have seen some bad videos that make simplistic descriptions of the diffusion rendering method that imply that you can completely reverse render every dataset from any model. With a small experimental dataset/model sure why not, but with a practical one with a massive properly curated and trained dataset, it is impossible. Again, only things that are overtrained by having too many duplicates in the source dataset will exhibit this. Models that can do this are essentially flawed.

    When you finetune SD using dreambooth with your own single image you need to crank up the training steps so much that it bleeds into other concepts in order to be legible, at which point you have effectively broken the model. To say that every image that it is trained on can be reproduced is simply not true, unless you make a bad, impractical and unusable model to specifically support this premise.

    It will depend on how unique a piece of input data is (most are not that different from another one if the dataset is huge) if it can be recreated accurately or not. It's amazing how close they can get to the originals if the correct coordinates in the model map is found, as close as a compressed jpeg is to an uncompressed image, but it will depend on the kind of image or data being replicated.
    Is it similar to sampling? Yeah I'd call it mega-sampling, more sampling than a human can ever do in a short space of time. It's purpose in mega-sampling is to find the correlations and biases in the dataset. In the human created image datasets it reveals the human image biases like a map. It's very easy to see that this mega-sampling is similar to normal sampling but on an industrial massive scale if the model is big, it's an infographic of the human input data. So unless using all data which you've input yourself, when "generating" material with these models it's much closer to curating than creating.
    If I try to create something in my mind using these models it never comes out how I want, it's always biased towards the models patterns, because I'm exploring a dataset. If it comes out how you want, then it's confirming your biases align with the model.

    Yah it definitely does reveal trends and bias. If you do want to create something that is in your mind the best method is to feed generations back in via image to image (with something like Automatic1111) and maybe using inpainting/outpainting. Better yet, mashing outputs together in photoshop and feeding them back in to re-synergize them.

    I still wouldn't equate it to sampling, more like physical modeling synthesis, if it could be created by feeding in recorded sounds that are then deconstructed into synthesis and able to be morphed together etc. Sampling as a metaphor just feels too coarse, like saying collage.

  • supadomsupadom

    @NeuM said:

    @supadom said:

    @echoopera said:

    @supadom said:
    @echoopera You make comparison to sampling. Well, it’s a spectrum. There’s those who mush 2 songs together DJ style and call themselves producers. There’s other who use the sampling form in a more subtle way.
    Who is the judge here?

    I’m equating it to how i use it. I can only judge myself and talk about my process. I have no control over what others do or think on the matter.

    But i do have my opinions of course 🤪

    Absolutely, I’m not pointing fingers at you at all, or at nobody in particular, in fact. I’m just musing about dilution of skill, of which (skill) there’s plenty of around, no doubt. Although I find it rather like the fake news situation where it’s hard to tell one thing from another.

    I’m not saying that IA cannot be used positively, just it confuses the shit out of me as it seems to be breaking some well established hierarchies. Maybe that’s what progress is supposed to do, I don’t know…and maybe that’s how people’s art takes the revenge on the privileged. As I said, I’m well confused by it. Especially when my friends some of whom are decent songwriters tell me how they’re using IA in times of creative drought.

    In school, the everybody gets a medal always rubbed me the wrong way but it sure has its merits.

    Frankly, it doesn't surprise me that your songwriter friends should use these tools. They're tools. Artists and musicians are typically the first to adopt new things, and I see no difference with machine learning tools.

    Folks used to be inspired by sunsets, social inequality etc, etc…now we get inspired instead of saying: if stuff it’s not coming out maybe I should take a break or go travelling. Now they press a button and voilá!

    Not my style, but don’t let that stop you from reaching for the skies.

  • NeuMNeuM

    @supadom said:

    @NeuM said:

    @supadom said:

    @echoopera said:

    @supadom said:
    @echoopera You make comparison to sampling. Well, it’s a spectrum. There’s those who mush 2 songs together DJ style and call themselves producers. There’s other who use the sampling form in a more subtle way.
    Who is the judge here?

    I’m equating it to how i use it. I can only judge myself and talk about my process. I have no control over what others do or think on the matter.

    But i do have my opinions of course 🤪

    Absolutely, I’m not pointing fingers at you at all, or at nobody in particular, in fact. I’m just musing about dilution of skill, of which (skill) there’s plenty of around, no doubt. Although I find it rather like the fake news situation where it’s hard to tell one thing from another.

    I’m not saying that IA cannot be used positively, just it confuses the shit out of me as it seems to be breaking some well established hierarchies. Maybe that’s what progress is supposed to do, I don’t know…and maybe that’s how people’s art takes the revenge on the privileged. As I said, I’m well confused by it. Especially when my friends some of whom are decent songwriters tell me how they’re using IA in times of creative drought.

    In school, the everybody gets a medal always rubbed me the wrong way but it sure has its merits.

    Frankly, it doesn't surprise me that your songwriter friends should use these tools. They're tools. Artists and musicians are typically the first to adopt new things, and I see no difference with machine learning tools.

    Folks used to be inspired by sunsets, social inequality etc, etc…now we get inspired instead of saying: if stuff it’s not coming out maybe I should take a break or go travelling. Now they press a button and voilá!

    Not my style, but don’t let that stop you from reaching for the skies.

    I get your point, but the genie is out of the bottle already.

  • Danny_MammyDanny_Mammy
    edited February 2023

    As long as you pay for and reference the original artist material its ok as at some point the tech will be able to locate the source material and copyright strike you. It has always been that way.
    I see artists using AI in conjunction with their own source material creating interesting stuff, however using the source material from other artists via AI to bang out a picture is consumption..

  • NeuMNeuM

    …> @Danny_Mammy said:

    As long as you pay for and reference the original artist material its ok as at some point the tech will be able to locate the source material and copyright strike you. It has always been that way.
    I see artists using AI in conjunction with their own source material creating interesting stuff, however using the source material from other artists via AI to bang out a picture is consumption..

    As I mentioned further up the thread, I think it’ll just come down to licensing at some point.

  • supadomsupadom

    @NeuM said:
    …> @Danny_Mammy said:

    As long as you pay for and reference the original artist material its ok as at some point the tech will be able to locate the source material and copyright strike you. It has always been that way.
    I see artists using AI in conjunction with their own source material creating interesting stuff, however using the source material from other artists via AI to bang out a picture is consumption..

    As I mentioned further up the thread, I think it’ll just come down to licensing at some point.

    The whole copyright has gone so bad with grooves being recognised as proprietary like in the case of Pharell Williams I cannot see it having a positive effect on music. Feels very big brother but I guess that tech revolution for ya.

    @NeuM said:

    @supadom said:

    @NeuM said:

    @supadom said:

    @echoopera said:

    @supadom said:
    @echoopera You make comparison to sampling. Well, it’s a spectrum. There’s those who mush 2 songs together DJ style and call themselves producers. There’s other who use the sampling form in a more subtle way.
    Who is the judge here?

    I’m equating it to how i use it. I can only judge myself and talk about my process. I have no control over what others do or think on the matter.

    But i do have my opinions of course 🤪

    Absolutely, I’m not pointing fingers at you at all, or at nobody in particular, in fact. I’m just musing about dilution of skill, of which (skill) there’s plenty of around, no doubt. Although I find it rather like the fake news situation where it’s hard to tell one thing from another.

    I’m not saying that IA cannot be used positively, just it confuses the shit out of me as it seems to be breaking some well established hierarchies. Maybe that’s what progress is supposed to do, I don’t know…and maybe that’s how people’s art takes the revenge on the privileged. As I said, I’m well confused by it. Especially when my friends some of whom are decent songwriters tell me how they’re using IA in times of creative drought.

    In school, the everybody gets a medal always rubbed me the wrong way but it sure has its merits.

    Frankly, it doesn't surprise me that your songwriter friends should use these tools. They're tools. Artists and musicians are typically the first to adopt new things, and I see no difference with machine learning tools.

    Folks used to be inspired by sunsets, social inequality etc, etc…now we get inspired instead of saying: if stuff it’s not coming out maybe I should take a break or go travelling. Now they press a button and voilá!

    Not my style, but don’t let that stop you from reaching for the skies.

    I get your point, but the genie is out of the bottle already.

    Agreed.

  • CarnbotCarnbot

    @AudioGus said:

    @Carnbot said:

    @AudioGus said:

    @Carnbot said:

    @AudioGus said:

    @Carnbot said:

    @AudioGus said:

    @Danny_Mammy said:
    Just like audio sampling one day they will be able to trace the original content the AI scanned and copyright claim the F%^k out of everyone. Be prepared.

    Diffusion rendering is nothing like audio sampling. Not to say that it has totally ethical roots of course, and while there are a few rare cases of overfitting in even the best ML models (where you can tell the source material) for the most part in 99.99999% of diffusion renderings things are completely, effectively laundered into oblivion.

    That's not strictly true, the data is in the model and can be replicated and reconstructed quite accurately but it's never asked to usually or it happens accidentally, since you have very little control over the data with text to image/audio apps. So while it's not exactly like traditional sampling it's a different type of sampling, eg data sampling, but it's closer to it than not.

    The images cannot be replicated and reconstructed unless they are overfitted, for example with many extraneous duplicates present in the original dataset when the model was trained. In the case of Stable Diffusion it was trained on billions of images and the model is only 4gigs. There is no way you can replicate and reconstruct those billions of images from a 4 gig file. There are certainly some examples of overfitting there but the vast overwhelming majority are vapor.

    The data is compressed with other data where there are similarities, so it won't come out exactly the same, but quite close. that's what it's doing really, removing the data it doesn't need from the masses of images to replicate a similar pattern and reduce the model, so it's creating relationships between all the data it's analysed. I think that looking at it like a data visualisation explorer is a good perspective as shown in the video above.

    I have seen some bad videos that make simplistic descriptions of the diffusion rendering method that imply that you can completely reverse render every dataset from any model. With a small experimental dataset/model sure why not, but with a practical one with a massive properly curated and trained dataset, it is impossible. Again, only things that are overtrained by having too many duplicates in the source dataset will exhibit this. Models that can do this are essentially flawed.

    When you finetune SD using dreambooth with your own single image you need to crank up the training steps so much that it bleeds into other concepts in order to be legible, at which point you have effectively broken the model. To say that every image that it is trained on can be reproduced is simply not true, unless you make a bad, impractical and unusable model to specifically support this premise.

    It will depend on how unique a piece of input data is (most are not that different from another one if the dataset is huge) if it can be recreated accurately or not. It's amazing how close they can get to the originals if the correct coordinates in the model map is found, as close as a compressed jpeg is to an uncompressed image, but it will depend on the kind of image or data being replicated.
    Is it similar to sampling? Yeah I'd call it mega-sampling, more sampling than a human can ever do in a short space of time. It's purpose in mega-sampling is to find the correlations and biases in the dataset. In the human created image datasets it reveals the human image biases like a map. It's very easy to see that this mega-sampling is similar to normal sampling but on an industrial massive scale if the model is big, it's an infographic of the human input data. So unless using all data which you've input yourself, when "generating" material with these models it's much closer to curating than creating.
    If I try to create something in my mind using these models it never comes out how I want, it's always biased towards the models patterns, because I'm exploring a dataset. If it comes out how you want, then it's confirming your biases align with the model.

    Yah it definitely does reveal trends and bias. If you do want to create something that is in your mind the best method is to feed generations back in via image to image (with something like Automatic1111) and maybe using inpainting/outpainting. Better yet, mashing outputs together in photoshop and feeding them back in to re-synergize them.

    I still wouldn't equate it to sampling, more like physical modeling synthesis, if it could be created by feeding in recorded sounds that are then deconstructed into synthesis and able to be morphed together etc. Sampling as a metaphor just feels too coarse, like saying collage.

    I would go further in saying that not only does it find biases, that's exactly what it's designed to do. These tools are scientific data analysis tools which have been re-branded as art "generators". In fact the same algorithms and tools are used in surveillance capture and analysis. Human work is surveyed en-masse and reduced and compressed to data, which is very cool and interesting but very disturbing without consent. Yeah if you transform the output significantly then you are introducing a process into the work, and art is a process, so you've done something at least. But that's true of any artmaking process.

    So while I agree that maybe it needs a different terminology in the creative world of the term "sampling". In the scientific idea of the word sampling is exactly what it is doing, taking and analysing samples of human creativity data and feeding it back to you.

  • AudioGusAudioGus

    @Carnbot said:

    @AudioGus said:

    @Carnbot said:

    @AudioGus said:

    @Carnbot said:

    @AudioGus said:

    @Carnbot said:

    @AudioGus said:

    @Danny_Mammy said:
    Just like audio sampling one day they will be able to trace the original content the AI scanned and copyright claim the F%^k out of everyone. Be prepared.

    Diffusion rendering is nothing like audio sampling. Not to say that it has totally ethical roots of course, and while there are a few rare cases of overfitting in even the best ML models (where you can tell the source material) for the most part in 99.99999% of diffusion renderings things are completely, effectively laundered into oblivion.

    That's not strictly true, the data is in the model and can be replicated and reconstructed quite accurately but it's never asked to usually or it happens accidentally, since you have very little control over the data with text to image/audio apps. So while it's not exactly like traditional sampling it's a different type of sampling, eg data sampling, but it's closer to it than not.

    The images cannot be replicated and reconstructed unless they are overfitted, for example with many extraneous duplicates present in the original dataset when the model was trained. In the case of Stable Diffusion it was trained on billions of images and the model is only 4gigs. There is no way you can replicate and reconstruct those billions of images from a 4 gig file. There are certainly some examples of overfitting there but the vast overwhelming majority are vapor.

    The data is compressed with other data where there are similarities, so it won't come out exactly the same, but quite close. that's what it's doing really, removing the data it doesn't need from the masses of images to replicate a similar pattern and reduce the model, so it's creating relationships between all the data it's analysed. I think that looking at it like a data visualisation explorer is a good perspective as shown in the video above.

    I have seen some bad videos that make simplistic descriptions of the diffusion rendering method that imply that you can completely reverse render every dataset from any model. With a small experimental dataset/model sure why not, but with a practical one with a massive properly curated and trained dataset, it is impossible. Again, only things that are overtrained by having too many duplicates in the source dataset will exhibit this. Models that can do this are essentially flawed.

    When you finetune SD using dreambooth with your own single image you need to crank up the training steps so much that it bleeds into other concepts in order to be legible, at which point you have effectively broken the model. To say that every image that it is trained on can be reproduced is simply not true, unless you make a bad, impractical and unusable model to specifically support this premise.

    It will depend on how unique a piece of input data is (most are not that different from another one if the dataset is huge) if it can be recreated accurately or not. It's amazing how close they can get to the originals if the correct coordinates in the model map is found, as close as a compressed jpeg is to an uncompressed image, but it will depend on the kind of image or data being replicated.
    Is it similar to sampling? Yeah I'd call it mega-sampling, more sampling than a human can ever do in a short space of time. It's purpose in mega-sampling is to find the correlations and biases in the dataset. In the human created image datasets it reveals the human image biases like a map. It's very easy to see that this mega-sampling is similar to normal sampling but on an industrial massive scale if the model is big, it's an infographic of the human input data. So unless using all data which you've input yourself, when "generating" material with these models it's much closer to curating than creating.
    If I try to create something in my mind using these models it never comes out how I want, it's always biased towards the models patterns, because I'm exploring a dataset. If it comes out how you want, then it's confirming your biases align with the model.

    Yah it definitely does reveal trends and bias. If you do want to create something that is in your mind the best method is to feed generations back in via image to image (with something like Automatic1111) and maybe using inpainting/outpainting. Better yet, mashing outputs together in photoshop and feeding them back in to re-synergize them.

    I still wouldn't equate it to sampling, more like physical modeling synthesis, if it could be created by feeding in recorded sounds that are then deconstructed into synthesis and able to be morphed together etc. Sampling as a metaphor just feels too coarse, like saying collage.

    I would go further in saying that not only does it find biases, that's exactly what it's designed to do. These tools are scientific data analysis tools which have been re-branded as art "generators". In fact the same algorithms and tools are used in surveillance capture and analysis. Human work is surveyed en-masse and reduced and compressed to data, which is very cool and interesting but very disturbing without consent. Yeah if you transform the output significantly then you are introducing a process into the work, and art is a process, so you've done something at least. But that's true of any artmaking process.

    So while I agree that maybe it needs a different terminology in the creative world of the term "sampling". In the scientific idea of the word sampling is exactly what it is doing, taking and analysing samples of human creativity data and feeding it back to you.

    Gotcha, being a music forum I just assumed that kind of sampling, heh

  • echooperaechoopera

    Madness:

  • AlexYAlexY

    I've been using some free AI plugins with the cut-up technique as an input for my own personal use. Been trying to get horror themed stuff. It's hit and miss, but sometimes it's good

  • echooperaechoopera

    Stuff is definitely happening out there:
    https://www.theverge.com/2023/2/6/23587393/ai-art-copyright-lawsuit-getty-images-stable-diffusion

    And it needs to happen :)

  • Danny_MammyDanny_Mammy

    and so, it begins..

  • echooperaechoopera
    edited February 2023

    @Danny_Mammy said:
    and so, it begins..


    If I were a CourtRoom artist, I'd use AI to render all the images I drew during the case proceedings. :) Wouldn't that just be the coolest meta move ever!! :)

  • echooperaechoopera

    Another day, another couple hours layering and having fun:

  • AudioGusAudioGus
    edited February 2023


    Popcorn indeed

  • NeuMNeuM
    edited February 2023

    @AudioGus said:

    Popcorn indeed

    Like I said, this is about money and these vast repositories of art and imagery will want to be compensated via licensing for the use of their property. That's all. Same goes for artists and illustrators. They should work out deals and then these art tools would make the original artists money and be able to provide customers with the best possible foundation for creating new works of art.

  • AudioGusAudioGus
    edited February 2023

    @NeuM said:

    @AudioGus said:

    >

    Popcorn indeed

    Like I said, this is about money and these vast repositories of art and imagery will want to be compensated via licensing for the use of their property. That's all. Same goes for artists and illustrators. They should work out deals and then these art tools would make the original artists money and be able to provide customers with the best possible foundation for creating new works of art.

    Thing is with the Richard Prince one in 2013 buddy just slapped a cutout of a guitar on someone else's photo and added some dots on the face and they actually won in court. Looks a lot more like the original to me then the Getty Ball mutation. But yah maybe the big olde web scrape is a different thing then collage boy.

    https://www.artnews.com/art-in-america/features/richard-prince-wins-major-victory-in-landmark-copyright-suit-59404/

Sign In or Register to comment.

© Vanilla Keystone Theme 2024