Video uploaded to the neural network – Kommersant FM
[ad_1]
ChatGPT developers showed a neural network that creates videos based on description. The model, called Sora, is capable of generating high-quality videos up to a minute long. Similar services have already appeared, but the videos were produced with severe distortion. The peculiarity of the new development is precisely in the level of detail. The neural network can create photorealistic videos and animate static images.
The developers claim that the model understands not just user requests, but also how generated objects exist in the real world. Artificial intelligence specialist Igor Shcherbakov discusses the prospects of the project: “It seems possible to use it as a cheap content generator. In principle, the creation of some kind of virtual brand ambassadors is already in trend.
Now that a neuron has appeared that generates, albeit short, but quite high-quality, believable videos, this could become a kind of start to the race for digital content, not only in the form of photographs and text, but also in the form of graphic video.
The model is now available for testing. But there don’t seem to be any big obstacles to the fact that in the future, and quite close at that, it will begin to generate longer videos and some kind of content, including in video games. This will be a step towards improving the quality and believability of pictures in the gaming industry.
At the same time, I think that this tool will become an assistant for those who work with images. It was exactly the same after the advent of large language models: it seemed that now many would begin to lose their jobs, then AI-based tools appeared that work well with pictures, and now with video. Essentially, this is a change in the toolkit of specialists in this field.
That is, if previously someone drew it with their hands, now they have to choose the right prompt and have experience working with a neuron. This is a technological revolution that will not only deprive people of work, it will change its format and set of tools.”
OpenAI admits that Sora does not work perfectly: the model may not be able to cope with creating a complex scene. For example, a person can take a bite of a cookie, but it will remain intact. In addition, it is necessary to take into account two serious risks associated with the development of such models, noted MTS Red specialist Alexander Baulin: “There are already at least two complaints about artificial intelligence. The first is related to copyright infringement. The systems are trained on databases, and if they contain copyrighted content, there is a chance of reproducing it. That is, AI developers either must agree on the use of these pictures, or take into account that in no case should they generate something from the training base.
And the second thing is that if we generate images that look like real videos, then we run the risk of scammers using deepfakes.
In this regard, all tools can serve for good, or they can cause harm. For example, a hammer can be used to hammer a nail, or it can be used to break a window for some evil purpose.
In the case of a neural network working with video, the fundamental difference is that usually deepfakes do not show the face well from different angles, that is, if the character turns his head back and forth, then it does not turn out so well. And here the background is generated, that is, it is a single picture. If Sora is as accessible as ChatGPT, which I’m not sure it will be, then it might be a more accessible tool.”
Sora is currently undergoing closed testing; experts are looking for potential critical problems in it. It is not yet known when the public version will appear.
Everything is clear with us – Telegram channel “Kommersant FM”.
[ad_2]
Source link
