By this logic, you can copy a copyrighted imege as long as you decrease the resolution, because the new image does not contain all the information in the original one.
Just because something is defined legally instead of technologically, that doesn’t make it vague. The modification violates copyright when the result is a derivative work; no more, no less.
The issue with this definition is that it’s overly broad. For instance, a hash of a picture could not exist without that picture. Nor do certain downscalings, like 2x2, 3x3 or 4x4. There must be an exact pixel value you can legally downscale any image to without violating copyright. Similarly, there is a point where creating a book’s synopsis starts violating copyright and where a song sounds too similar to another one.
And based on their size, LLMs - in my opinion - cannot possibly violate copyright for their source material because they couldn’t possibly store more than a couple of bits per work. Only works that occue frequently in the training data can actually be somewhat reproduced by LLMs.
By the way, fair use doesn’t even exist in every - including my - jurisdiction.
This has lead to people being successfully sued for copyright infringement because they posted pictures of their home online that contained a copyrighted wallpaper in the background.
In the case of Stable Diffusion, they used 5 billion images to train a model 1.83 gigabytes in size. So if you reduce a copyrighted image to 3 bits (not bytes - bits), then yeah, I think you’re probably pretty safe.
Your calculation is assuming that the input images are statistically independent, which is certainly not the case (otherwise the model would be useless for generating new images)
By this logic, you can copy a copyrighted imege as long as you decrease the resolution, because the new image does not contain all the information in the original one.
More like reduce it to a handful of vectors that get merged with other vectors.
Am I allowed to take a copyrighted image, decrease its size to 1x1 pixels and publish it? What about 2x2?
It’s very much not clear when a modification violates copyright because copyright is extremely vague to begin with.
Just because something is defined legally instead of technologically, that doesn’t make it vague. The modification violates copyright when the result is a derivative work; no more, no less.
What is a derivative work though? That’s again extremely vague and has been subject to countless lawsuits seeking to determine the bounds.
If your work depends on the original, such that it could not exist without it, it’s derivative.
I can easily create a pixel of any arbitrary color, so it’s sufficiently transformative that it’s considered a separate work.
The four fair use tests are pretty reliable in making a determination.
The issue with this definition is that it’s overly broad. For instance, a hash of a picture could not exist without that picture. Nor do certain downscalings, like 2x2, 3x3 or 4x4. There must be an exact pixel value you can legally downscale any image to without violating copyright. Similarly, there is a point where creating a book’s synopsis starts violating copyright and where a song sounds too similar to another one.
And based on their size, LLMs - in my opinion - cannot possibly violate copyright for their source material because they couldn’t possibly store more than a couple of bits per work. Only works that occue frequently in the training data can actually be somewhat reproduced by LLMs.
By the way, fair use doesn’t even exist in every - including my - jurisdiction.
This has lead to people being successfully sued for copyright infringement because they posted pictures of their home online that contained a copyrighted wallpaper in the background.
In the case of Stable Diffusion, they used 5 billion images to train a model 1.83 gigabytes in size. So if you reduce a copyrighted image to 3 bits (not bytes - bits), then yeah, I think you’re probably pretty safe.
Your calculation is assuming that the input images are statistically independent, which is certainly not the case (otherwise the model would be useless for generating new images)