There’s usually nothing left of the original image. But sometimes a specific image pops up in the dataset more often and gets overtrained, which is why you can get a pretty close copy of the Starry Night from vanilla SD. But yeah, it’s not tracing.
Legitimately, it’s like these people have no understanding of the actual technology.
The other response you’ve received take about a very small subset of obtained images, which makes sense for being replicated, but anyone who trained on creating a specific image a million times would be able to replicate that image easily. Even then it takes a lot of luck and effort to accurately replicate the exact image.
If you are not specifically trying to recreate an overly popular image, then there is practically no element left from any particular image.
Considering that it is effectively acting on a pareidolia interpretation of static represented by countless possible prompt and setting combinations, the copyright issue should only really be relevant when people use the tool specifically trying to recreate a particular work. Literally any other paint program would make that easier.
As an artist, in regards to the pareidolia aspect, I do virtually the same thing when illustrating an image. Disney/Warner can already afford as many peasants to learn whatever styles they want. I can’t afford a team of lackeys. I can however use an open source diffusion model to create entirely unique and personally tailored and designed illustrations.
Existing concept of copywrite does not work, and if people should argue anything, it should be that wealthy businesses specifically have much more restriction and responsibility.
I’m personally excited for a future where peasant artists can also create complex beautiful works using these tools.
That’s not true at all. AI uses latent noise as a medium to draw images, there’s nothing left of the original image in its dataset.
There’s usually nothing left of the original image. But sometimes a specific image pops up in the dataset more often and gets overtrained, which is why you can get a pretty close copy of the Starry Night from vanilla SD. But yeah, it’s not tracing.
Legitimately, it’s like these people have no understanding of the actual technology.
The other response you’ve received take about a very small subset of obtained images, which makes sense for being replicated, but anyone who trained on creating a specific image a million times would be able to replicate that image easily. Even then it takes a lot of luck and effort to accurately replicate the exact image.
If you are not specifically trying to recreate an overly popular image, then there is practically no element left from any particular image.
Considering that it is effectively acting on a pareidolia interpretation of static represented by countless possible prompt and setting combinations, the copyright issue should only really be relevant when people use the tool specifically trying to recreate a particular work. Literally any other paint program would make that easier.
As an artist, in regards to the pareidolia aspect, I do virtually the same thing when illustrating an image. Disney/Warner can already afford as many peasants to learn whatever styles they want. I can’t afford a team of lackeys. I can however use an open source diffusion model to create entirely unique and personally tailored and designed illustrations.
Existing concept of copywrite does not work, and if people should argue anything, it should be that wealthy businesses specifically have much more restriction and responsibility.
I’m personally excited for a future where peasant artists can also create complex beautiful works using these tools.