OpenAI's Sora (Text-to-Video model)

[foreverlurk]

I don't know if you have seen OpenAI's latest demo of their new Text-to-Video generative AI, but right now it's blowing my mind. I'm not sure how flexible or censored this new model will be - I hope us mortals will be granted access one day. Even crazier, it also simulates physics, "Sora is a data-driven physics engine. It is a simulation of many worlds, real or fantastical. The simulator learns intricate rendering, "intuitive" physics, long-horizon reasoning, and semantic grounding, all by some denoising and gradient maths."

https://openai.com/sora

More samples with prompts:
https://twitter.com/OpenAI/status/1758192957386342435

Minecraft simulation demo:
https://twitter.com/andrewwhite01/statu ... 2909474137

I also found a sizey-ish prompt on Twitter, featuring two miniature pirate ships fighting in a cup of coffee:
https://twitter.com/DrJimFan/status/1758210245799920123

I don't want to overhype it, I can't help but imagine if generating SW videos will be possible...

[Raso719]

Oh man I wonder how much processing power is needed to do something like this. I can't imagine the trial will be as simple as with ChatGPT...

[foreverlurk]

I know, right now it's being tested by a few privilieged industry pros. I'm not sure it'll be available at large. They also mentionned they are implementing "safety" features (ie. heavy censorhip) so I'm not holding my breath. If it's the same as DALL-E, we're only going to be able to use a fraction of its true power, because "safety".

"Open" AI, my arse.

[shrinker_s]

I was thinking that the opportunities of this are great. However, a lot of the "powerful" AI's have been locked down to make them "safe" -- However I have seen a few AI pages allowing more "Adult" content. I have also seen some pretty spicy stuff posted on the Dark Web. Maybe in a year or so there will be a site for us "More Creative" types to explore.

[foreverlurk]

Oh it will be censored, no doubt about it. I still want to try it to generate sizey content, and maybe try to find a way around the filters like we were able to do previously with image gen.

Meanwhile... the model seems to understand size difference pretty well. Just think of the possibilities:

Giant duck in the streets of Boston
https://twitter.com/_tim_brooks/status/ ... 4974760042

Miniature train
https://twitter.com/model_mechanic/stat ... 0956219476

Can't wait to try it!