Live data from Hacker News

Ask HN: DALL-E was trained on watermarked stock images?

news.ycombinator.com

41–50 of 233 posts

Re: Ask HN: DALL-E was trained on watermarked stock images?

#41

These are the absolute worst DALL-E images I've seen. Do people generally just share the amazing ones and most of the output is actually complete shite? Like Instagram presenting the top 1% of people's lives.

Of course people are more likely to share the best iamges – or in this case, the one most illustrative of their concern (about watermarks).

Also: my sense is that getting the best results often requires a lot of extra coaching with style/detail words. As we can't see the prompt here, we don't know what sort of style/details were requested. GIGO.

Re: Ask HN: DALL-E was trained on watermarked stock images?

#42
You may want to use the native 'Share' option, especially on the one with the watermark.

You'll get a public link, at `labs.openai.com` rather than some random image-sharing site, which will show the image & the prompt used to generate it (including a credit to "your-first-name × DALL·E").

Re: Ask HN: DALL-E was trained on watermarked stock images?

#43
post #41

These are the absolute worst DALL-E images I've seen. Do people generally just share the amazing ones and most of the output is actually complete shite? Like Instagram presenting the top 1% of people's lives.

Of course people are more likely to share the best iamges – or in this case, the one most illustrative of their concern (about watermarks). Also: my sense is that getting the best results often requires a lot of extra coaching with style/detail words. As we can't see the prompt here, we don't know what sort of style/details were requested. GIGO.

You're right. This shows the prompt and it doesn't have such style directives

https://ibb.co/gz5RDkB

Re: Ask HN: DALL-E was trained on watermarked stock images?

#44

These are the absolute worst DALL-E images I've seen. Do people generally just share the amazing ones and most of the output is actually complete shite? Like Instagram presenting the top 1% of people's lives.

Top 1% is a bit exaggerated, but there is definitely a lot of not good stuff. I find that Dall-E does especially poorly with underspecified prompts too, unlike something like Midjourney which can give visually pleasing photos for even the most abstract concepts. Dall-E tends to do better with concrete and specific prompts.

Here's an example: Stressful Shapes

Dall-E: https://i.imgur.com/JBkSh0y.png

Midjourney: https://i.imgur.com/C02Zq3i.png

On the other hand, here's a specific prompt: "nerdy yellow duck reading a magical book full of spells"

Dall-E: https://i.imgur.com/FMKZ8zc.png

Midjourney: https://i.imgur.com/lpsg6af.png

Re: Ask HN: DALL-E was trained on watermarked stock images?

#45

> but surely you can't just... use stock photos without paying for the license? They aren't hosting the infringing content. Training on the data is probably covered under fair use. Generations are of _learned_ representations of the dataset, not the dataset itself. This makes it closer to outputting original works (probably owned by the person who used the model). The players involved here are known for being litigio…

What if I write a machine learning algorithm that only generates images that it has seen in the training dataset, with one pixel slightly different.

Re: Ask HN: DALL-E was trained on watermarked stock images?

#46
post #40

All large-scale public machine learning stuff is depending on being exempt from copyright restrictions, under fair use doctrine. Look at my responses to all of the threads about Copilot + GPL for more info about that application of it: https://hn.algolia.com/?query=chrismorgan+copilot+gpl&type=c... . When that is finally tried in court, if it fails to any meaningful extent at all (including going all the way up to Su…

Great points but scary. If training ML models on copyrighted data becomes illegal in the US but remains legal in say China or Russia then the US will quickly fall Behind on ML capabilities - major national security implications at the very least. I suspect if the decision went the way you suggest congress would have to change the law to allow training.

Isn't that true for all technology? In the U.S. we have the specwriter system which leads to inefficiency to get around copyright. In China or Russia they just copy the code and iterate.

Re: Ask HN: DALL-E was trained on watermarked stock images?

#47

What is interesting is a human analogy. Say you were an artist who went to every art show and museum and studied all the art there. If you produced a work of art solely from memory that contained large portions of other people's copyrighted art, would that still fall under copyright/require licensing?

There are definitely lines to be crossed. Let me tell you about one of them.

There is a comics creator named Kieth Giffen. He's done a lot of solid work over the years for DC and Marvel, there's a playful love of the medium and its history that flows through a lot of his work. At first his style was pretty middling; nothing terrible, nothing to really stand out from the pack. Then one day his work changed dramatically - he got a lot more daring in spotting his blacks, inking with a heavier brush, and doing a lot of panels that were a closeup of a backlit head with rim lighting, and eyes and teeth standing out in white. It was grounded in observation but had a lot of fresh ways to abstract a scene in the service of story. It was like nothing else on the racks and really striking.

It was also completely swiped from the work of an Argentinian artist named José Muñoz. Pick up one of Muñoz's shadow-drenched crime stories, put it next to one of Giffen's superhero tales, and you could clearly see the influence. And not just the influence, influence is okay - Giffen had started entirely cloning Muñoz's style, completely dropping all his other influences in the process. Muñoz was not happy when he heard about this, and neither were other artists in the field of comics. Influence is one thing, everyone's influenced by other artists, and if you're familiar with an artist's influences you can tell. But dropping all your other influences to start drawing almost exactly like a new one? That's just not done.

Giffen got a lot of shit for this. Giffen quit comics for a couple of years after this, and when he came back he had a new look. He still does the Shadowy Muñoz Face now and then but it's more along the lines of one of the many things he's borrowed from his multiple influences rather than one of the ways he was wholesale ripping off Muñoz.

"Style theft" is completely legal in the eyes of the court. There was nothing legally actionable going on here. But in the court of his fellow artists, Giffen was judged, and found guilty.

There's a range here. Nobody's going to care if you pick up a collection of Winsdor McCay's pioneering 19xx comic strip "Little Nemo" and do a dream-themed story that borrows his distinctive panel composition, lettering, and inking choices. Nobody's going to care if you do one drawing that precisely lifts Mike Mignola's heavy use of black and thin, clear lines. If you do superheros long enough then you're pretty much obligated to do at least one story that emulates Jack Kirby as closely as you can. If you worked as someone's assistant for a half a decade then you are very much allowed to bust out a perfect rendition of their style at any point in your entire life. But there is definitely a line you can cross where every artist (and a lot of non-artists) who sees a side-by-side view of what you're doing and what you're swiping from will say "dude, not cool, stop swiping their style".

These image generators actively encourage adding the names of prominent, living artists to your prompts to get the results you want. Is this crossing the same line Kieth Giffen did?

Re: Ask HN: DALL-E was trained on watermarked stock images?

#48
post #30

I am not a lawyer, but I've had to argue about copyright with several. In the United States, there are two bits of case law that are widely cited and relevant: In Kelly v. Arriba Soft Corp (9th), found that making thumbnails of images for use in a search engine was sufficiently "transformative" that it was ok. Another case, Perfect 10 (9th), found that thumbnails for image search and cached pages were also transforma…

Search engines don't create market harm for a work because they don't compete with it. In fact, they do the opposite: they advertise the work, making it more accessible and increasing exposure. These AI tools on the other hand seem to do the exact opposite. They can (or could, if they got good enough) absolutely compete with a work, and therefore seem like they create substantial market harm. The character of use als…

As I see it, 3 of the 4 tests are strongly in OpenAI's favor; the 'market effect' is mixed.

(1) The use is highly transformative;

(2) the images used were offered to the anonymous browsing public (with watermarks);

(3) the end effect of training will only retain a tiny spectral distilled essence of any individual photo, or even a giant source corpus;

(4) there's a potential risk of market competition from the ultimate model output, for some uses – but that's also the most 'transformative' aspect.

Getty et al could potentially just ask creators of such models not to include their images – perhaps by blocking their crawling 'User-Agent' – and it might not make any real difference in the models.

Post reply on HN