OpenAI Launches Sora Video Generator (axios.com) 30
ChatGPT maker OpenAI released its AI-generated video tool called Sora for general use by its paying customers Monday. From a report: The company then said it would do wide testing with creatives and red-teaming with security experts before its release to the public. "We don't want the world to just be text," OpenAI CEO Sam Altman said in a live-streamed announcement Monday.
"[Video] is important to our culture," Altman added. The company said in a statement that the latest version of Sora, which will be offered as a standalone product to ChatGPT Plus and Pro customers, is "significantly faster" than the one it previewed. It lets you generate videos up to 20 seconds long.
"[Video] is important to our culture," Altman added. The company said in a statement that the latest version of Sora, which will be offered as a standalone product to ChatGPT Plus and Pro customers, is "significantly faster" than the one it previewed. It lets you generate videos up to 20 seconds long.
The mistakes are hillarious! (Score:2)
While the generated videos are impressive, I chuckle at the mistakes I've seen immediately, such as:
- in the Chinatown scene, how many people morph from walking away from the camera to walking towards the camera
- where that last bear cub appears from in the nature video
- the tiny colour-changing car parked on the front lawn in the 1950s style video
I can see a 'Damn you autocorrect'-type meme site evolving from these types of goofs.
Scenes are here (Score:4, Informative)
Since there's no direct links in the post or article, the above mentioned videos are here on their main site [sora.com].
Cheers (Score:2)
The mistakes are absolutely hillarious! (Score:2)
While the generated videos are impressive, I chuckle at the mistakes I've seen immediately, such as:
- in the Chinatown scene, how many people morph from walking away from the camera to walking towards the camera
The video examples are here [openai.com].
Taking one in particular, the alien walking on a busy street, note that at the very beginning, over the alien's right shoulder, the taxicab sliding sideways on the road. almost at the end a taxicab slides dangerously across the crosswalk through the crowd of people.
Everyone is walking in the street.
People disappear as a cab passes in front of them. (Behind the Alien to his right, two people walking across the street and not "down" the street")
Hilarious! OpenAI should totally anno
Re: (Score:2)
Several of them are very usable and could be aired as a TV style advertisement. The bird/Buddha can be a signature for a travelling agency; the puppy in superman costume can be used for a toy store (or the furry thing). When there's too many characters it gets confused. Just don't use it for that use case, or extract short sequences that do not exhibit the problem. If this tool does with little effort what used to require weeks of planning and cost many thousand in street production and CGI, it has a great
Re: (Score:2)
Several of them are very usable and could be aired as a TV style advertisement. The bird/Buddha can be a signature for a travelling agency; the puppy in superman costume can be used for a toy store (or the furry thing). When there's too many characters it gets confused. Just don't use it for that use case, or extract short sequences that do not exhibit the problem. If this tool does with little effort what used to require weeks of planning and cost many thousand in street production and CGI, it has a great future.
Truthfully, if it's cheap enough it won't matter how badly it garbles the end-game. It'll get used.
Re: (Score:2)
Re: The mistakes are hillarious! (Score:1)
Re: (Score:2)
Archive them. The stuff is so funny, but a few years from now the generators will get as good as the current generation of image generators and we will lose the fun aspect of the AI misunderstanding the scenes and producing hilarious videos.
Re: (Score:2)
Like with most things AI (art, music, etc), it generates things that are amazing but commonly contain minor flaws, which people then become hypersensitive to. You then have two choices:
1) Be a lazy arse and just post it as-is.
2) Use actual editing skills to turn it from nearly-perfect to perfect.
#2 often involves going back and forth between the AI model and traditional editing software.
Re: (Score:2)
When it comes to video editing, apart from regenerating bad segments (or even individual objects in the scene), you can do an awful lot with even basic things like pan/crop and zoom, global blurs, mask blurs, etc (blurs can be used to mimic changes in camera focus), cutting off the start or end of segments or adding jump cuts into long, continuous segments, etc.
Just me? (Score:5, Insightful)
"[Video] is important to our culture," Altman added.
Anybody else have to fight a huge gag reflex at Altman trying to discuss the importance of anything to our culture? Some vulture like him even thinking the term "culture" means he's most likely trying to find a way to completely subvert it for his own profit. Yuck.
Re: (Score:2)
I would caution about mixing fake AI generated video so freely with the foundational content, it will just lead to upheaval.
The end of influencers (Score:2)
Who needs needs those narcissists with their sponsored posts bragging about their daily cup of tea when you can have Sora generate your daily dose of ass and titties tailored just for you with a few keywords?
Re: (Score:2)
That seems to be the main usage of open-source AI image tools these days :P
Re: (Score:2)
Replace the narcissists with corporate sponsored AI! BRILLIANT!
Great, more videos I do not want to see... (Score:3, Interesting)
Give me some well-written text instead any day. Of course Artificial Ignorance cannot do that either.
Re: (Score:3)
Give me some well-written text instead any day. Of course Artificial Ignorance cannot do that either.
Hence the "Arrogant Ignorance" of A.I.
Re: (Score:2)
That is a new one for me. Fits well!
Re: (Score:1)
...Slashdotted :D
Those days are long gone, especially now that Slashdot is but a crypto scene news outlet [slashdot.org] disguised as a 'tech news site'.
I doubt a C64 webserver would even hiccup [reddit.com] with twice or triple Slashdot's current daily traffic.
Re: (Score:1)
...Slashdotted :D
Those days are long gone, especially now that Slashdot is but a crypto scene news outlet [slashdot.org] disguised as a 'tech news site'.
I doubt a C64 webserver would even hiccup [reddit.com] with twice or triple Slashdot's current daily traffic.
Right; that's why I said "in the old days, I would have said ... "
The age of automatic mediocrity is upon us (Score:4, Informative)
Now talentless human content creators aren't even needed to create poor content anymore.
Re: (Score:1)
Unlikely - usually there's a seed that is set, but it can be made to be 'random' so the same prompt will generate similar but fairly different output (say like telling 10 people to simply "draw a flower" will result in likely 10 very different pictures of flowers of differing colours and type)
To get something close to exactly the same (if really possible) usually requires duplicating several parameters (including the base model used) in addition to the prompt.
Re: (Score:2)
Diffusion models are initialized with random noise, which is gradually denoised with a bias toward the text prompt. If the guidance toward the prompt is not very strong (e.g. "Mona Lisa" will definitely guide you to a very specific image) the outcome is mostly random while still being faithful to the prompt. Also there are a lot of parameters, schedulers, step sizes, etc. that have influence on the outcome.