>>69592514History lesson:
Generative image AI emerged organically from research by NVIDIA on dynamic scaling algorithms. This research formed the basis of all those funny image filters you find built into phone photo apps and photo-centric special media like instagram, and evolved from the shitty upscale techniques used to butcher old media for digital rereleases in the early 2010s.
The goal was to find a way to algorithmically predict what should be in a pixel based on existing information present in an image, so that the algorithm could, for example, take a 100x100 pixel image and blow it up to 1000x1000 without losing too much quality. This "superesolution" method allows you to render something in a lower resolution (which takes less processing power) and then use the algorithm (which is cheap to process) to upscale it and fill in the blanks left by adding 990^2 more pixels.
To accomplish this, it needs to determine 2 things: the location of pixels that need editing, and the colour the pixel needs to be. Both of these things can be quantified and then reduced to a mathematical formula. Upscaling algorithms would find this information via a tonne of metadata baked into the image made specifically to quantify values for this formula to use. This tech has been "mature" for almost a decade, but has improved as the assumptions the algorithm has to make about what should be in a pixel has been refined, both by better, more efficient data and by some meaningful improvements in the underlying mathematics.
Generative image AIs (they aren't really AI but we'll get to this) pair this algorithm, which is already a mature technology, with an image recognition Neural Network. NNs are probably well understood at this point so I won't belabour the point, but they basically let the developers set an initial series of criteria and then the training process guides the model in creating the quantifiable tags for image attributes themselves. In other words, it creates the metadata we talked about earlier via training instead of needing it provided externally. The rest of the process is exhaustively well understood and not in any way new. Your graphics card is doing it as we speak. What makes these AI models "unique" is just that training and refining neural network until it stops spitting back out eldritch horror is expensive and time-consuming process, which is essentially all that these various AI companies have actually done. But as is already widely understood, there are serious ethical issues with using someone else's intellectual property as training data without their express permission or any form of compensation--which is one reason that these companies clutch pearls over keeping their training data confidential
>>69593619>these discussions will peter out into inevitability.Actually this kind of generative AI is on borrowed time. Adversarial imaging is on the horizon (it's technically already here, check your captcha) and when that becomes widely available to end-users rights holders will be able to foil NN piracy with a simple watermark. While in an ideal world this would create a more equitable market where people are compensated fairly for their work, I suspect it will just expose the whole industry as a grift that can't be profitable if it has to pay for things it was taking for free.
