Hacker News Reader

Comment by sorenjan | Hacker News Reader

As I was doomscrolling Reddit yesterday I saw the first AI video that fooled me[0]. It's a very typical Reddit video that's unremarkable in a lot of ways, yet designed to perfectly fit in and attract engagement from a large part of the users. I guess you can say it's "too perfectly calibrated, suspiciously optimized to trigger maximum relatability" just like the post in the article. On one hand I don't care that much since I already dislike these kinds of videos even when they're "real", on the other hand the amount of slop is about to increase even more. The best time to quit Reddit was years ago, the second best is today.

[0] https://www.reddit.com/r/SipsTea/comments/1jxnau3/does_size_...

View on HN

I'm not sure in what way that is AI generated. I'm following this space quite a bit and it looks like it's at least partially real. The voice is very realistic, I doubt that's AI generated. The motion and facial expressions are also very natural. I would assume it's some AI transformation (like style transfer) on top of a real video. I see no evidence that it's AI. In fact, I think there's a genre of online garbage that's claiming that things were made by certain AI tools when they weren't. I'm not saying this video has no AI in it, just that it doesn't seem fully AI and we have no way to tell.

Fingers

I don't believe video generation can make nonverbal communication sync up so well, regarding the shrug, eye movement, facial expression etc. perfectly synced with the voice. As I said, I think it's conditioned on some real footage, somewhat like ControlNet perhaps.

Look at the handbags hanging in the air next to people walking by in the background, and the six fingers on the interviewer's hand. But yes, I had to look at it repeatedly to see that it's fake, it's definitely good enough to fool most people scrolling their feed.

I don't follow it super close, but I've seen some videos from the Wan model that's now available in ComfyUI, and they can be really good.

I suspect it is a ControlNet style generation, driven by some real underlying footage. The nonverbal communication is just too good for what I've seen so far from pure AI generation.

Maybe someone took a cropped video and used some AI video model to do out painting. I've found a few different versions of it using Google lens, some are vertical and show more of the image vertically but are more cropped horizontally, some have an Invideo AI water mark on them, but at different positions. I can't find any video on Invideo AI's page that is close to this kind of facial expressions and lip sync, so I also think at least the face is real video.

Apparently this is the original video: https://www.instagram.com/reel/DAYWr0HoWmN/

Like you said, most of is not AI, my best guess is that somebody used AI to do inpainting to remove the captions and watermark. First time I've seen anything like it.

And now it’s gone, removed by mods

Reply on news.ycombinator.com