Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Show HN: Fake or real? Try our AI image detector

trial.nuanced.dev

Post reply on HN
Print view View original

21–30 of 52 posts

Re: Show HN: Fake or real? Try our AI image detector

#21

Post by john_noumenonic »

jprete wrote: post #5
htrp wrote: post #2

Cool hack to get some Human Feedbacked data. Unfortunately your system doesn't seem to be able to upload an image. https://trial.nuanced.dev/demo/upload_progress has an event stream that polls every 2 seconds or so but doesn't seem to return any success criterion.

Data collection for adversarial training was also my first thought. The same training data used for classifying images as AI-generated can also be used to train the AI to generate images that fool more people.

True enough, though we're interested in stress testing our model to see where the gaps are right now.

Re: Show HN: Fake or real? Try our AI image detector

#22

Post by koito17 »

GaggiX wrote: post #18

I like how confident the model is when you provide an anime image, the model always thinks it's AI even though I only provided images created by humans, I don't think I've ever seen a worse AI detector in my life. I hope that this is not a real demo of the product.

Heh, I tried the exact same thing. I uploaded sketches from my favorite artists. Specifically, sketches produced between 2012 and 2016. All of them were identified as AI with greater than 50% probability.

Of course, if one uploads recent sketches, one could be cynical and claim the artist traced over AI-generated image. But I have never seen this done in practice

Re: Show HN: Fake or real? Try our AI image detector

#24

Post by john_noumenonic »

simlevesque wrote: post #4

I tried with this eclipse photo: https://www.cnet.com/a/img/resize/535a36e2cb72f06e9b3dc04254... It said 92% AI. Do you have any stats about how often it gets it right ?

We're generally seeing an accuracy of 90% on test sets, with its distribution primarily being images generated by diffusion models and the "real" images that are more ground level imagery. We'll have to take astrophotography into account!

Re: Show HN: Fake or real? Try our AI image detector

#26

Post by evbogue »

egypturnash wrote: post #17
evbogue wrote: post #7

Earlier quoted context omitted.

AI == An Illustrator!? You got me thinking down the path that sometimes when people say AI they mean something very different than what I thought AI meant.

Adobe Illustrator. I edited the comment to make that clear. :)

lol. Now I will know what people are talking about at all the cool tech parties.

Re: Show HN: Fake or real? Try our AI image detector

#27

Post by rfrey »

GaggiX wrote: post #18

I like how confident the model is when you provide an anime image, the model always thinks it's AI even though I only provided images created by humans, I don't think I've ever seen a worse AI detector in my life. I hope that this is not a real demo of the product.

> don't think I've ever seen a worse AI detector in my life. I hope that this is not a real demo of the product.

"I tried it on anime images and it didn't work well on that class" would have been sufficient.

Re: Show HN: Fake or real? Try our AI image detector

#30

Post by bee_rider »

I think iPhones do some pretty complex image processing, and other brands of phones go even farther. I believe there have been instances of phones adding in completely new details (although maybe I hallucinated that detail).

What’s the expected result and also what should we put as a “true” answer if we take a picture with our phones and upload it?

Post reply on HN