For a few months now, I have been trying the new tools of AI art to generate images from text prompt (Midjourney, Dalle2). It appears that image generating AI has big difficulties to generate hands and texts. I know some of the reality checks in Lucid dreaming are about looking at your hands or reading a text to see if something is wrong.
Do you think that the image generation that are difficult for the AI neural network are also difficult for the brain in lucid/dream state ? If you have some knowledge in AI art generation, is there any other similarities you would have noted ?
AI art is trained on existing artwork. This is a major ethical debate because artists have not consented to their work being used in this way for all AI models (specifically Lensa/Stable Diffusion is guilty of this). Basically, the AI is using pieces of existing artwork (and even things like sexual content that was produced without consent to all parties involved) to produce art. I haven’t heard about Midjourney and Dalle2, so I can’t speak on those AI specifically and do not know if they have ethically sourced the images used to train the AI.
Given that, I would suspect and suggest the that hands and texts being difficult for the AI to render may possibly have to do with the artwork the AI is trained on. Texts in artwork are often artistically manipulated and hands are notoriously difficult to draw for many artists.
I am not super informed on the technical aspects of the AI art generation process so this may not be the case, but it was my first thought with what I’ve read and seen about AI art.
I do think there’s a place for AI art in the future that can be ethically sourced so that artists get compensated and credited for their contributions. It could be a fantastic tool and fun experience for artists and non-artists alike, but currently there are a lot of ethical concerns about art theft and consent with the current AIs being used and I would encourage you to look into the discussion on this as there are problematic aspects to using AI to generate art.
Well, you can choose on what sources you train the AI. So you can make an ethical AI that’s trained only on work that’s old enough to be in public domain. I hope someone will make it soon.
About hands and text, I don’t know. You can have good looking texts, it’s just that the letters are not in the right order.
I actually was wondering about this topic myself not so long ago. I think maybe there are similarities in a way that we as humans are not used to looking at hands for a very long time or recreating something so tidy and neat as texts. Just like the ai is not trained enough (yet) to do so, but can only create things it has focussed most on recreating and leaves out details it considers not as important (just like us in dreams)