So it's really good, and we have reason to believe, never again, anything that happens in a video. Unless there's a super-product somewhere to authenticate footage?
At first usage I'm not impressed. I've probably spent a couple grand on Seedance 2 to date, and I can't find anything google omni flash does better than Seedance from running a handful of samples through the system. You can find some of the videos I've made in my HN bio link.
I'm an AI optimist. But AI video is probably the one thing that does depress me. Seeing that we can make anything visually, there's nothing that impresses me visually. I watch a video that two years ago I would've thought was really cool, and now my first thought is, "Yawn, is this AI?".
Video, more than anything else, is the place where I really care if something is AI or not. If I could get a TikTok that had no AI usage -- I'd be in. Which is weird for me, because I'm typically the guy who is all-in on AI.
I think it is like around 2010 or so I use to upload just this god awful music to early Sound Cloud because it was easy to make music with a DAW.
I even remember being on a psytrance production music mailing list 25 years ago and 95% of the tracks people posted were absolutely terrible, including myself.
I have seen a few incredible pieces from AI video but most has just not been that interesting. Then even the incredible pieces are 5 second one offs. No narrative, no continuity. I think of a random, real 5 second clip from Clockwork Orange with no backstory or context in the movie, who cares? Even the most visually interesting scenes wouldn't make sense and would be boring.
Right now it seems like we are at the stage of sampling random 5 second clips from early sound cloud and concluding this is the artistic utility of an entire new technology like DAW software and VST synths. That is obviously absurd.
In my day job I program rigid body behaviour in real time amongst other simulations.
I think rigid body contact is hard to learn as it is inherently discontinuous.. something you discover when trying to code a solver.
As such I always use this prompt as a test:
"A video of a jenga brick tower falling over as a brick is removed. The physics of each brick must be realistic."
It gave me a video of where bricks suddenly disapper or morph into others[1]. The linked video is after 2-3 iterations of me insisting on realistic physics. If you are just glancing at this, you would believe it is realistic.
That said this is still very impressive and one more step towards .. IDK what. But I am a bit reasurred that at least my job won't be fully replaced with AI :)
Classic 3d simulation artifact with boundary conditions. I remember for an assignment where I had to model liquid with rigid bodies, they would suddenly gain infinite force at the corner and just disappear. It's clear that they must have used a lot of these kinds of synthetic data. But what's impressive to me, every release of these models, I am feeling less and less uncanny valley.
I'm not sure why, especially because you're a developer... But damn, the amount of people that expect AI to just one shot stuff is hilarious. Half of the time I make a typo or something, should I be laughed out of the room?
Does anyone else feel like Google is just always a dollar short and a day late here? Maybe not a dollar short, but it's like they've consistently been focused on the wrong thing. First they missed chatbots, now they're missing coding agents while they double down on chatbots and video gen (which OpenAI has already basically abandoned). Maybe this strategy is actually genius and I'm too stupid to grasp it.
Nano Banana Pro is still the industry standard as far as I’m concerned. I think giving a vision model spatial awareness is the next evolutionary step here, so I don’t think they’re behind at all.
While at a cursory glance it looks as impressive as always, subtle spatial errors, and geometry that changes as it goes out of sight and comes back again hints at the fact that Google has still yet to solve the problem of deep spatial understanding.
Which considering just how pretty and detailed this whole thing looks, imo points at a fundamental issue at how these things are trained - it's as if there's no structure to its knowledge and training, like how an artist trained to draw would first try to understand simple 2d composition, then perspective, then light and shadow, mastering each concept and gradually building up a hierarchical understanding - it seems like its trying to learn everything at once.
I would rather see an AI model that I could give a floorplan of a building and it would generate an accurate flythrough on any path, even if it looked like butt.
Im not just talking out of my arse, I did work for a while in data science/engineering, and one of the big lessons people needed to be reminded of is to clean/downsample the data - a dataset consisting of a million samples could very well take 1000x as long to process as if we downsampled the whole thing to just a couple of thousand samples and we could learn the same conclusions with the fraction of expended time/effort.
I'm sure there's a similar logic in RL, that if you dump a trillion samples into the datacenter that consumes the same power as a city, what the model learns is what it could've learned with a much more curated training set and directed approaches.
It's funny how they specifically use the phrase "output that follows real-world physics" to describe the marble rolling video. At the end of the zigzag track, the marble jumps up for no reason. In a couple of other places it speeds up with no apparent energy source. It's still an amazing result, but they could have picked a better example for this claim!
What's the end goal of video generation? It feels unnecessary. Text generation leads to AI that can replace workers. Video generation is bad and only for video content generation, like movie and tv show production?
To be honest, I think the performance of Gemini Omni Flash is still not as good as Seedance 2.0. You can try using both models on this platform. https://omnivideoai.co
33 comments
[ 2.5 ms ] story [ 46.6 ms ] threadmodel card: https://deepmind.google/models/model-cards/gemini-omni-flash...
I did not create any videos yet.
Google, building great AI that nobody can try out.
But thx for the press release.
Certainly not me - you have to be a great artist /designer to even imagine what to do with it.
There's got to be a reason this is phrased so insanely, right?
Oh god...
Video, more than anything else, is the place where I really care if something is AI or not. If I could get a TikTok that had no AI usage -- I'd be in. Which is weird for me, because I'm typically the guy who is all-in on AI.
I even remember being on a psytrance production music mailing list 25 years ago and 95% of the tracks people posted were absolutely terrible, including myself.
I have seen a few incredible pieces from AI video but most has just not been that interesting. Then even the incredible pieces are 5 second one offs. No narrative, no continuity. I think of a random, real 5 second clip from Clockwork Orange with no backstory or context in the movie, who cares? Even the most visually interesting scenes wouldn't make sense and would be boring.
Right now it seems like we are at the stage of sampling random 5 second clips from early sound cloud and concluding this is the artistic utility of an entire new technology like DAW software and VST synths. That is obviously absurd.
As such I always use this prompt as a test: "A video of a jenga brick tower falling over as a brick is removed. The physics of each brick must be realistic."
It gave me a video of where bricks suddenly disapper or morph into others[1]. The linked video is after 2-3 iterations of me insisting on realistic physics. If you are just glancing at this, you would believe it is realistic.
That said this is still very impressive and one more step towards .. IDK what. But I am a bit reasurred that at least my job won't be fully replaced with AI :)
[1] https://streamable.com/2em1r3
Which considering just how pretty and detailed this whole thing looks, imo points at a fundamental issue at how these things are trained - it's as if there's no structure to its knowledge and training, like how an artist trained to draw would first try to understand simple 2d composition, then perspective, then light and shadow, mastering each concept and gradually building up a hierarchical understanding - it seems like its trying to learn everything at once.
I would rather see an AI model that I could give a floorplan of a building and it would generate an accurate flythrough on any path, even if it looked like butt.
Im not just talking out of my arse, I did work for a while in data science/engineering, and one of the big lessons people needed to be reminded of is to clean/downsample the data - a dataset consisting of a million samples could very well take 1000x as long to process as if we downsampled the whole thing to just a couple of thousand samples and we could learn the same conclusions with the fraction of expended time/effort.
I'm sure there's a similar logic in RL, that if you dump a trillion samples into the datacenter that consumes the same power as a city, what the model learns is what it could've learned with a much more curated training set and directed approaches.
It could make the comments section even more fun.
From a technical perspective, it's very impressive, no doubt. But from an artistic perspective I thought all of these examples on the site look bad.
I have not used Gemini in a month.