3 ms·

This really isn't going to change much. DALLE-2 has and will likely for a long time have a lot of imperfections that will never be fixed. Due to the inherit way

This really isn't going to change much. DALLE-2 has and will likely for a long time have a lot of imperfections that will never be fixed. Due to the inherit way that these GANs work it's not going to force anyone really out of business. At most it could help artists with their day to day tasks but not really much more. You in effect have to create the proper datasets to even generate proper images itself. Even if the sample size required is getting substantially lower, increasing the barrier to entry. It will probably take hundred of thousands of images at the least to create a decent looking image. Each of which mind you is copyrighted. The quality of the images will be a new requirement for the field. As Machine Learning Models are intellectual property we will likely see the unionification of artists as they work towards protecting the industry. Honestly the majority of these algorithms are going in the complete wrong direction and companies that use this will likely go bankrupt if google follows anything similar to amazon's business model.

ShamelessC 4y ago
> DALLE-2 has and will likely for a long time have a lot of imperfections Sure. > that will never be fixed. Less sure. > Due to the inherit way that these GANs Not a GAN. DALLE-2 uses a diffusion model for generation and a transformer/diffusion to convert CLIP text embeds to CLIP image embeds. DALLE-1 used a variational autoencoder, not a GAN. I mention this simply because both diffusion and VAE's seem to be more robust than GAN's (at the cost of slower inference times). > It will probably take hundred of thousands of images at the least to create a decent looking image. A little confused about this. DALLE-2 took hundreds of _millions_ of images to train, but it doesn't need even one to generate an image now. Having said that, finetuning requires small datasets like this and is perhaps what you meant? > Each of which mind you is copyrighted. I hear you on this - as it's a little devastating not to be even attributed for your work if you're a hard-working artist. However, training on datasets that aren't explicitly public domain or an open license is not a violation of copyright. Nor, at least presently, is releasing the weights of that model to the public. It would only be a violation to distribute the media itself - a problem which has been trivially circumvented by just releasing the URL's of the dataset for other researchers to download. In the case of smaller datasets, releases come with disclaimers that the data is not to be used for anything other than research. OpenAI of course has not even released the URL's for their web datasets. > As Machine Learning Models are intellectual property Hm, I think I agree that this could become an issue. You can see this in the licensing terms in StyleGAN2/3 already, for instance (typical nvidia, they basically own your outputs last time I read the thing). At least in terms of what I've been seeing lately however, this won't be a very effective regulatory moat for AI companies. Research tends to be replicated rapidly by lots of brilliant open source developers. Sometimes, the paper even comes out long before the model is released giving plenty of time for a clean-room implementation. There are also of course great institutions like in Heidelberg doing bleeding edge, state of the art work and releasing _everything_. See e.g. latent-diffusion. Machine learning requires lots of hardware - that's the best moat these companies can have. NVIDIA GPU's and Google TPU's are what's to be concerned about here.
- yike321123 4y ago
  > Machine learning requires lots of hardware - that's the best moat these companies can have. NVIDIA GPU's and Google TPU's are what's to be concerned about here. With each concurrent generation the cost to produce the tools necessary with decrease exponentially as hardware also become more freely available. I don't see hardware as a limiting factor. Though the two monopolies have been known to work together to slow down technological progress. Honestly I think that our progress would likely have gone a lot faster without them. But for some strange reason they would rather follow moore's law. I don't really believe moore's law exists, there's limits to everything and the rules they mention are unsustainable and unrealistic. It makes me wonder where the real end of it is or if we've made the wrong decisions along the way solely for profit motive. How many times has funding been cut for research programs simply because nothing happened for a single quarter? Like most Generative Artificial Networks, DALLE-2 is not really anything new. In the computer graphics field several research groups have been discussing similar techniques for decades. I've even seen a few demos by various colleges of their own implementations. The idea that DALLE is the future next step for several fields such as art is misleading at best and straight up delusional at worst. The main reason I state the issues within DALLE-2 not being fixed is mainly due to these factors. It's the wrong direction for this type of problem especially so for the billions of parameters that they use. The hyperfixation that google is attempting to do to this industry for it's own financial interests will pigeonhole several companies towards suboptimal solutions that will at best give them a somewhat workable solution at an expensive price. This of which will likely bankrupt them in the process as even with improvements to techniques will still cost them in the millions to produce even a few of these samples simply due to the labor cost involved. While the art industry has been transformed by the technology industry through the gig economy intentionally for this purpose. Even at those rates you still end up with the cost of producing the models being outcompeted by the value provided. This is especially true as art improves overtime, the level of detail and fidelity along with varying art styles. Not to mention the fact it is an unsustainable model that prey on artists by systemically abusing them by turning them into slaves. But who am I say that it won't just collapse. After all companies such as walt disney, pixar, and dreamworks have been known for doing just this for decades. As other countries build "incentive" plans to systemically enslave their workers as well to outsource the art work towards studios abroad as well this likely won't change. Even the best animation studios such as Studio Ghibli only pay just above the poverty line. While it's never been my dream to draw or even write the idea that people of any kind being subject to abuse is something that I can't bear to see. All I can say is my deep hate for this never ending rat race. I always find it interesting when I read about these algorithms how the people who read, learn and create them seemingly have no sense of empathy for those that they seem to see as beneath them. Though I'm not claiming to say that you, yourself are to blame or are even contributing towards it. It's just something that bothers me. Anyways, continuing on towards copyright. It's interesting to me how these datasets are generated is through the theft of artwork. I've already mentioned the labor cost involved in generating their own datasets and to mention it's unlikely anyone besides google will be able to afford to get away with this simply due to the magnitude of the lawsuit they would face. (It would likely be class action.) It always makes me wonder just how detrimental such as case would be and the effects of it due to how the U.S. Judicial System works. Though Silicon Valley has always been known for it's corruption and bribing of government officials, who knows? They might win. But it doesn't necessarily mean it will affect the entire industry as a whole but it might. It's interesting because there are already several papers on this matter that show that these neural networks work by copying those images. But you'll never know how a company like google can spin that. I always found it interesting that programmers think of machine learning as a black box. The tool itself is largely unaffordable for large scale implementation and you would be better off working on other GAN's instead to help artists. Of course there's no one size fits all solution to this problem even though google would like to claim otherwise the quality of the images are useless for professionally done products without requiring heavy altercations. Teaching an entire generation to rely on this tool is likely going to end up counterproductive especially if it results in them losing the fine skills that are required to be developed for the task. I can see this going in several ways that could all end up horribly such as younger artists being forced to draw for the models themselves or being forced to be turned in essentially a data encoder writing labels for the dataset without having an practical skills or learning much. Or in how it's currently going the work being forced into sweatshop like conditions in 3rd world countries with an English speaking majority. Worsening the existing caste system within this industry. Maybe we'll see snip bits of it like on the tags of our jeans and tee shirts secretly hidden away, when a member of the upper caste clicks to generate an image instead what's returned are the words help me no food no water instead. The respective employee fired after attribution of the code is uncovered. Or on a lighter note artists secretly drawing penises on everything in order for the neural network to secretly generate penises on every image it draws. I'm frankly upset. Not at you but an industry. Sorry if this sounds like a tangent or if it sound like I'm blowing up on you. While the ideas mentioned may sound like my own they are not. While I'm unable to give sources to whose they are but I stand behind them.

ShamelessC 4y ago

> DALLE-2 has and will likely for a long time have a lot of imperfections Sure. > that will never be fixed. Less sure. > Due to the inherit way that these GANs Not a GAN. DALLE-2 uses a diffusion model for generation and a transformer/diffusion to convert CLIP text embeds to CLIP image embeds. DALLE-1 used a variational autoencoder, not a GAN. I mention this simply because both diffusion and VAE's seem to be more robust than GAN's (at the cost of slower inference times). > It will probably take hundred of thousands of images at the least to create a decent looking image. A little confused about this. DALLE-2 took hundreds of _millions_ of images to train, but it doesn't need even one to generate an image now. Having said that, finetuning requires small datasets like this and is perhaps what you meant? > Each of which mind you is copyrighted. I hear you on this - as it's a little devastating not to be even attributed for your work if you're a hard-working artist. However, training on datasets that aren't explicitly public domain or an open license is not a violation of copyright. Nor, at least presently, is releasing the weights of that model to the public. It would only be a violation to distribute the media itself - a problem which has been trivially circumvented by just releasing the URL's of the dataset for other researchers to download. In the case of smaller datasets, releases come with disclaimers that the data is not to be used for anything other than research. OpenAI of course has not even released the URL's for their web datasets. > As Machine Learning Models are intellectual property Hm, I think I agree that this could become an issue. You can see this in the licensing terms in StyleGAN2/3 already, for instance (typical nvidia, they basically own your outputs last time I read the thing). At least in terms of what I've been seeing lately however, this won't be a very effective regulatory moat for AI companies. Research tends to be replicated rapidly by lots of brilliant open source developers. Sometimes, the paper even comes out long before the model is released giving plenty of time for a clean-room implementation. There are also of course great institutions like in Heidelberg doing bleeding edge, state of the art work and releasing _everything_. See e.g. latent-diffusion. Machine learning requires lots of hardware - that's the best moat these companies can have. NVIDIA GPU's and Google TPU's are what's to be concerned about here.

yike321123 4y ago
> Machine learning requires lots of hardware - that's the best moat these companies can have. NVIDIA GPU's and Google TPU's are what's to be concerned about here. With each concurrent generation the cost to produce the tools necessary with decrease exponentially as hardware also become more freely available. I don't see hardware as a limiting factor. Though the two monopolies have been known to work together to slow down technological progress. Honestly I think that our progress would likely have gone a lot faster without them. But for some strange reason they would rather follow moore's law. I don't really believe moore's law exists, there's limits to everything and the rules they mention are unsustainable and unrealistic. It makes me wonder where the real end of it is or if we've made the wrong decisions along the way solely for profit motive. How many times has funding been cut for research programs simply because nothing happened for a single quarter? Like most Generative Artificial Networks, DALLE-2 is not really anything new. In the computer graphics field several research groups have been discussing similar techniques for decades. I've even seen a few demos by various colleges of their own implementations. The idea that DALLE is the future next step for several fields such as art is misleading at best and straight up delusional at worst. The main reason I state the issues within DALLE-2 not being fixed is mainly due to these factors. It's the wrong direction for this type of problem especially so for the billions of parameters that they use. The hyperfixation that google is attempting to do to this industry for it's own financial interests will pigeonhole several companies towards suboptimal solutions that will at best give them a somewhat workable solution at an expensive price. This of which will likely bankrupt them in the process as even with improvements to techniques will still cost them in the millions to produce even a few of these samples simply due to the labor cost involved. While the art industry has been transformed by the technology industry through the gig economy intentionally for this purpose. Even at those rates you still end up with the cost of producing the models being outcompeted by the value provided. This is especially true as art improves overtime, the level of detail and fidelity along with varying art styles. Not to mention the fact it is an unsustainable model that prey on artists by systemically abusing them by turning them into slaves. But who am I say that it won't just collapse. After all companies such as walt disney, pixar, and dreamworks have been known for doing just this for decades. As other countries build "incentive" plans to systemically enslave their workers as well to outsource the art work towards studios abroad as well this likely won't change. Even the best animation studios such as Studio Ghibli only pay just above the poverty line. While it's never been my dream to draw or even write the idea that people of any kind being subject to abuse is something that I can't bear to see. All I can say is my deep hate for this never ending rat race. I always find it interesting when I read about these algorithms how the people who read, learn and create them seemingly have no sense of empathy for those that they seem to see as beneath them. Though I'm not claiming to say that you, yourself are to blame or are even contributing towards it. It's just something that bothers me. Anyways, continuing on towards copyright. It's interesting to me how these datasets are generated is through the theft of artwork. I've already mentioned the labor cost involved in generating their own datasets and to mention it's unlikely anyone besides google will be able to afford to get away with this simply due to the magnitude of the lawsuit they would face. (It would likely be class action.) It always makes me wonder just how detrimental such as case would be and the effects of it due to how the U.S. Judicial System works. Though Silicon Valley has always been known for it's corruption and bribing of government officials, who knows? They might win. But it doesn't necessarily mean it will affect the entire industry as a whole but it might. It's interesting because there are already several papers on this matter that show that these neural networks work by copying those images. But you'll never know how a company like google can spin that. I always found it interesting that programmers think of machine learning as a black box. The tool itself is largely unaffordable for large scale implementation and you would be better off working on other GAN's instead to help artists. Of course there's no one size fits all solution to this problem even though google would like to claim otherwise the quality of the images are useless for professionally done products without requiring heavy altercations. Teaching an entire generation to rely on this tool is likely going to end up counterproductive especially if it results in them losing the fine skills that are required to be developed for the task. I can see this going in several ways that could all end up horribly such as younger artists being forced to draw for the models themselves or being forced to be turned in essentially a data encoder writing labels for the dataset without having an practical skills or learning much. Or in how it's currently going the work being forced into sweatshop like conditions in 3rd world countries with an English speaking majority. Worsening the existing caste system within this industry. Maybe we'll see snip bits of it like on the tags of our jeans and tee shirts secretly hidden away, when a member of the upper caste clicks to generate an image instead what's returned are the words help me no food no water instead. The respective employee fired after attribution of the code is uncovered. Or on a lighter note artists secretly drawing penises on everything in order for the neural network to secretly generate penises on every image it draws. I'm frankly upset. Not at you but an industry. Sorry if this sounds like a tangent or if it sound like I'm blowing up on you. While the ideas mentioned may sound like my own they are not. While I'm unable to give sources to whose they are but I stand behind them.