(Un)fortunately every Sora video has the OpenAI watermark in the bottom right.
Exactly. Bit of a giveaway...
Even without the watermark, some of the videos look very real
At the end it just gives you the final score without telling which ones you got wrong, which is highly unsatisfactory.
The video goes over the answers and points out some tells from each.
Very bad mobile experience, any slightly missed click and the modal is gone and you are starting over.
I got 10/10. At least 3 of them I remembered from Sora's announcement page (or samas twitter thread the same day). 3 or 4 of them seemed obvious, just having a lot of experience with cameras personally, and having seen a lot of AI content. I just chose the video that looked like the biggest pain-in-the-ass to do for real. One of them looked like a f/0.9 lens. Ain't nobody got time for that irl. I got lucky on the rest, I wasn't sure.
Just pick the one with higher saturation.
And less reflections/movement.
That didn't work particularly well, which was one of my main takeaways from the quiz.
Most striking in the train trip videos, where the cityscape video with extensive reflections is the "fake".
Protip: Remove the watermarks first.
Way easier to add the watermark to the real ones, ha.
I didn't see any watermarks, I think they fixed it. It's so hard!
One still has it on there, you can see the vertical lines at the very bottom of the image, not cropped out yet.
does it just piece bits of videos together from the randomness of the internet to generate something? So it's a statistical match to some aggregate combination of videos?
It's not random. Most of the fake videos are taken from OpenAI's recent Sora demo.
I was wondering on the internals of Sora
It's a diffusion transformer architecture, so no it doesn't piece together pieces of video. If you're familiar with denoising algorithms, the diffusion algorithm is essentially a semantically guided denoising algorithm. If you feed it pure noise so the only information it has is the semantic guiding, it will generate a video from that noise directly. I'm not sure exactly how the transformer part of the algorithm contributes, but my guess is that it's giving the denoiser the ability to not just look at adjacent pixels in 2D space, but across time through the attention mechanism. That's just a guess, though.
Thank you, after reading your comment, I did some research and stumbled upon this, an explanation of how Sora works from Jim Fan:
Are you asking here how AI art works, in general? That would take more than fits in a comment, and there are lots of explainers online. You could even ask ChatGPT this question. But no, it's not piecing bits of videos together. It doesn't store enough of each individual piece of art to have a piece. The art might take up X bytes, X being a large chunk of the whole internet, but the AI is only a fraction of a percent of that. So it can't possibly store chunks of each piece of art. Even 1 pixel from each piece of art would be too much. But it does store the patterns the art had in common with each other. And from that, it generates.