Hacker News new | past | comments | ask | show | jobs | submit login

This is a common misconception. AI is not copying anything. It is studying the images in a similar way that humans do. Entire size of the stable diffusion model is around 6gb, with pruning it goes down to 3gb. Training set is in terabytes. So it is common sense that no copying is done.

I think artists need to be fair about AI, is there any artist that created their style without ever studying other artists? That is high improbable because humans need to observe to create art. There is even a saying that "Good Artists Copy; Great Artists Steal".




Still, people did not consent to their art to be trained on.

Just like how it is a bad idea to train github copilot on copyrighted code, or how the same company that made Stable Diffusion promised that they will not use copyrighted music for training (because they're scared of the music industry), copyrighted art should not be used on training sets without permission.

This is going to cause a lawsuit somewhere down the line, even if images only contribute a few bits each, signatures are still seen and this could make an argument in a court case.

It would be fair to everyone to only use public domain material for training sets.


Do people need to consent for other artists to use their work as references when painting? That seems pretty analogous to training an AI on a corpus of artwork.


It's generally not gone over well when it's discovered that an artist has been copying composition, poses, etc from pre-existing work of the same type. Manga and comic artists for example have had their careers derailed over it.

Stylistic mimicry is more murky, partially because humans aren't as good at cloning styles — inevitably the clone will take on some influence from the one doing the copying, which isn't true of AI. If you ask the AI to draw X in the style of Y artist, it's going to be a dead-on copy of Y artist's style.


Copying entire compositions is different from references. I'm case you're unfamiliar, artists usually have a few images off to the side of their workpiece to look at for inspiration and to reference colors and image elements.


That's the thing though, a it's not uncommon for ML-generated images have major elements that are almost 1:1 copies of existing pieces, especially if the prompt includes an artist's name or a style that's closely tied to a particular artist, going well beyond an artist using existing pieces for reference/inspiration.


Why do we make analogies to people? Machines have no rights of expression or inherent freedoms. There is no "learning", we're not "teaching" the machine anything. It boils down to heuristics and statistics. Imagine we're in the 1930s with no computers: If I were to study every every Agatha Christie book and write my own based on statistical likelihoods of words, characters, plot elements etc. I would be seen as a copy-cat hack-fraud author or worse. If I were to take inspiration of the crime/thriller/detective genre and write my own story in my own universe then I would be simply an author within the same genre.

ChatGPT/"ML" copy the fine details. We're getting artists signatures turn up in generated work... That's not inspiration, and I would argue not even transformative.


That is because the AI doesn't have enough context knowledge to distinguish prominent signatures from stylistic flourish or a thing being represented in the painting. The generated signatures themselves are usually very different, though in some cases they can come very close to the original artist signature when properly prompted for artists with "big" signatures. Better image part tagging will solve this.


If you want to get pendantic, the human brain ultimately functions off heuristics and inputs. Again, references are not copied, they are used to inspire and guide new artwork. Even artists producing totally original artwork use references.


The current ML approaches are not "studying", and have no thought process whatsoever, let alone "in a similar way that humans do". They build models that allow them to reproduce existing works in whole or in part—usually in many parts, put together to form a new work formed wholly out of those elements.

Humans learn actual techniques; they understand what the elements in their art actually mean; when they copy, they do so with intention (whether malicious or not). ML approaches to content creation are incapable of intention, because intention is the product of a conscious mind, and regardless of how similar some of the data structures involved in them are to certain models of the human brain, not one of the existing ML projects even remotely approaches anything we could term consciousness. (Nor is that even their purpose.)


Trained on 5 billion images to create a 6 GB model. Every image contributed no more than a few bits of information.


Billion and Giga is both 10^9, so about 6/5 or 1.2 byte or 10 bits, and somehow enough to regurgitate examples from training set[0]. Oh how convincing “AI is just learning as humans do” arguments are.

0: https://twitter.com/kortizart/status/1588915427018559490


Maybe this incredibly famous photo from a National Geographic cover is overrepresented in the training set?


Well I TIL humans are incapable of copying another piece of art even when asked to do so


It's never too late to learn to draw!


I could learn to mine pigments and create paints too but we have better options now


Pick up a pencil, start from drawing a circle, a real circular circle... Caveman in front of a VT100 is still a caveman. Caveman with a charred willow branch in hand, is something a bit more than that.


Your statement implies the source dataset's file sizes represent the amount of visual information in them. This is almost certainly not the case.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: