Do we need to nerf machine learning?
It’s fascinating to witness with a certain critical eye how the applications of a technology, machine learning, and its possibilities impact society.
Yesterday, Stable Diffusion announced a new version of its algorithm, 2.0, which, among other things such as preventing the generation of pornographic images or photorealistic portraits of celebrities, also eliminates the possibility of requesting drawings that copy the style of a particular artist. The restrictions have met with negative reactions from many users who feel that the model has been somehow nerfed, and that such restrictions limit their creative freedom.
In the image, the result of asking Stable Diffusion for “a bird in the style of Piet Mondrian”. Leaving aside that the image is rather basic and that I have not refined the request or retouched it in any way with the tools that the algorithm itself offers, the result is obvious, and corresponds to extracting from the library images created by Piet Mondrian, capturing his style, and superimposing on it the image of a bird, also obtained from thousands or millions of images of birds.
What is the problem, apart from starting, in this case, from the style of an instantly recognizable artist, but who has had one of his paintings hanging upside down in a gallery for seventy-five years without anyone noticing? Some artists…
