More than 20 years ago, I skipped class to watch The Matrix.
Morpheus said: Welcome to the real world.
I woke up from a dream, and when I returned to the classroom, the teacher was talking about the top-level design in the process of restructuring state-owned enterprises.
Is what you see real?
Time flies, and I finally forgot a lot of people, and finally forgot a lot of things.
When I woke up, the AI circle was boiling, and everyone found that reality no longer existed.
OpenAI released a teaser of a large model, capable of generating a 60-second complete.
This large model is called sora.
As a heavy user of AI tools, I use generative AI and AI painting tools for more than 2 hours a day on average.
One of the things that interests Sora the most is that the images it generates don't collapse or flicker.
In fact, at this stage, open source AI painting tools can also be generated**, but most of them are within 4 seconds, and the biggest drawback is face collapse and flickering.
The reason for the face collapse is that many large models are difficult to process the facial details of the large picture, and the reason for the flickering is that at this stage, the large models are generated frame by frame, and then put together **, and the details of each frame will flicker when there is a change.
From this point of view, in front of Sora, all the current ** large models have to kneel.
Some analysts believe that SORA uses a game engine. I agree with this point of view, using a game engine can cleverly bypass the drawbacks of generating frame-by-frame images and then stitching them together.
Jim Fan, a senior scientist at Nvidia, has some of his own thoughts on Sora:
SORA is a data-driven physics engine. It is a simulation of many worlds, both real and imaginary. The simulator learns complex rendering, intuitive physics, long-term reasoning, and semantic understanding through denoising and gradient learning.
I wouldn't be surprised if Sora was trained on a lot of synthetic data with Unreal Engine 5. It has to be!
In chatgpt35 Before the launch, ChatGPT released several versions, although the performance was amazing, but the ability was limited, and it was also tepid. 3.After 5 hit it popular, it made persistent efforts to launch a paid 4Version 0.
No matter how amazing it is, ChatGPT at least has a "prelude", its basic principles and operating logic, as well as progress, and people who are interested in AI have an understanding.
The emergence of SORA was a little unprepared, just yesterday, everyone was still racking their brains for the problem of ai** flashing, after all, everyone has seen the ** based on the existing level of technology**, knowing that the bottleneck is**, knowing that the upper limit is**. Sora is like jumping out of human technology, the sudden arrival of Zenith technology.
But! If you're using a game engine like Unreal Engine 5, all of this is understandable. It can only be said that the previous AI** took the old road of AI painting, and it is taken for granted that **= painting is superimposed frame by frame, and Sora cleverly changed the route, which is still the logic of ChatGPT, using data to drive the game engine, and then the game engine generates the picture.
Such technology does not surpass the current level of technology, it is just a bone surprise, which no one really imagined before.
ChatGPT first converts the received text into **, and then uses ** to drive the game engine to generate a picture.
With the success of SORA, more and more AI large models will be connected to the game engine in the future, which will cause dimensionality reduction to the existing large painting models!
As for the ** industry and the traditional film and television industry, it has a far-reaching impact.
This year's Spring Festival Gala Ren Suxi's song "She with Light on the Pillow", there is a scene of a ceramic villain dancing, which is painted with AI.
And it's with stable diffusion + controlnet + animatediff + lcm + ipadapter, based on sd15。
Fortunately, this year's Spring Festival came early!
If it is a few days later, with the rapid iterative growth of generative AI and game engine technology, even Ren Suxi himself can draw it with AI!
However, according to OpenAI's urine, there is a high probability that it will not be open source, but with the roadmap established, it is not difficult for other AI companies to quickly keep up.
The biggest pressure is probably Midjourney, which becomes more difficult to transform when generative AI + game engines become standard.
It can be said that film companies that don't make AI large models will have to die.