AI in underwater video

[Madmax7]

Pretty impressive (and I agree, scary). Seeing it after being told it's AI, and knowing a bit about what I'm looking at, I guess it's not that hard to pick some things out. It would definitely be easy to take it as real though, if I wasn't told and the video was of something I knew nothing about.

 

[Mobulai]

I had to do it about 10 times to get one where there weren't bubbles coming out after I had surfaced!

I wanna see the other 9

how much rebreather footage was in the training dataset

I still wonder if any generative AI has capability to differentiate OC from CC even on a graphical front — would it generate that much bubbles if you told it “descending diver” instead
I suspect there’s not enough labeling/annotation in the training data for that; so probably Flow just decided “it’s a good idea to make the diver move their hand by the mouthpiece” regardless

Still, for 2 pictures to start and 10 tries that’s nice outcome

 

[doctormike]

I wanna see the other 9

I still wonder if any generative AI has capability to differentiate OC from CC even on a graphical front — would it generate that much bubbles if you told it “descending diver” instead
I suspect there’s not enough labeling/annotation in the training data for that; so probably Flow just decided “it’s a good idea to make the diver move their hand by the mouthpiece” regardless

Still, for 2 pictures to start and 10 tries that’s nice outcome

I kind of doubt that there is that much information - distinguishing between OC and CC. The prompt was "generate a video of the diver surfacing", so it did understand that. If I tried the same thing with just a prompt that says "generate video", then it just does a cross fade between short clips made just from each image, with no real transition.

Here is another one that I generated from two photos. Note the bubbles on descent. Now CCR divers might notice that, and I could just say something like "oh, I had inadvertently overinflated my counterlungs and I was getting back to minimal loop volume", or "I was doing a dil flush".

 

[Mobulai]

Here is another one that I generated from two photos. Note the bubbles on descent. Now CCR divers might notice that, and I could just say something like "oh, I had inadvertently overinflated my counterlungs and I was getting back to minimal loop volume", or "I was doing a dil flush".

Exactly the part I was thinking about; my bet is if you give it 10 pictures/samples for a 40m descent on CCR, it will exhale bubbles continuously like that — as if it was OC; no amount of flushes can justify that.

Ofcourse I’m basing this guess on how images that get generated (from scratch) tend to be a twilight zone / uncanny valley mix of OC/CC and steampunk looking equipment

But the best usecase anyways is for simpler transitions/cut scenes like the one you shared originally

 

[doctormike]

Exactly the part I was thinking about; my bet is if you give it 10 pictures/samples for a 40m descent on CCR, it will exhale bubbles continuously like that — as if it was OC; no amount of flushes can justify that.

Ofcourse I’m basing this guess on how images that get generated (from scratch) tend to be a twilight zone / uncanny valley mix of OC/CC and steampunk looking equipment

But the best usecase anyways is for simpler transitions/cut scenes like the one you shared originally

Yeah, I think that in general, one of the problems that non-technical users have understanding AI is that they really do expect it to be "intelligent", and that becomes a semantic thing with questions about sentience, etc ..

True general AI is not really what these LLMs and video engines do, they just make guesses based on the statistical distribution of information in the training dataset. It's not like the system is learning about how rebreathers work to generate those videos, it is just modeling its output on the media diet it was fed.

Here's something that I made with a different system, where the AI engine put in so many bubbles, I had to make it into a joke.....

 

[mcohen1021]

Flow is a subscription based google service. They are currently offering 1-month for free, but I had little desire to plug in my credit card or paypal info to play with it just to have to figure out how to cancel it in 30 days.

 

[Mobulai]

Yeah, I think that in general, one of the problems that non-technical users have understanding AI is that they really do expect it to be "intelligent", and that becomes a semantic thing with questions about sentience, etc ..

True general AI is not really what these LLMs and video engines do, they just make guesses based on the statistical distribution of information in the training dataset. It's not like the system is learning about how rebreathers work to generate those videos, it is just modeling its output on the media diet it was fed.

100% on point

Here's something that I made with a different system, where the AI engine put in so many bubbles, I had to make it into a joke.....

The content my heart seeks — this should be the training video for all boom-drills