OpenAI Announces ‘Video ChatGPT’; material lasts one minute – 02/15/2024 – Tech

The creator of ChatGPT, OpenAI, announced this Thursday (15) the launch of a generative artificial intelligence model capable of generating videos of up to one minute from requests made in text.

The platform called Sora creates, according to OpenAI, complex scenes with multiple characters and different movements. Competitive models work with durations between four and ten seconds.

The new platform, for security reasons, still has access restricted to risk assessors and a limited number of visual artists and video producers. The company did not disclose a launch projection.

To avoid harmful use of fake videos known as deepfakes, already used to influence elections and carry out financial scams, OpenAI says it intends to add an encrypted seal to materials generated by Sora.

Furthermore, the company intends to replicate the abuse filter already active on ChatGPT on the new video generating platform. The company also claims that it has developed a second image classifier that will evaluate the security of each image.

The model, however, fails to maintain the coherence of physical effects, such as light and wind direction, and to reproduce cause and effect relationships. Spatial, right and left instructions, for example, also receive wrong answers.

In one case, researchers responsible for developing new generative AI show a video of a man running on a treadmill in reverse.

Sora was trained from a collection of videos to simulate real-world characteristics. Therefore, the name refers to the Japanese word for “sky”, which also means the “shell that envelops the world”.

To maintain the coherence of the videos for one minute, OpenAI adopted a strategy similar to that used in ChatGPT, which calculates what the next most likely word to appear in a text would be. Sora predicts the most likely next frames in a given sequence of images.

Sora, according to OpenAI, is a crucial step towards increasing the understanding of artificial intelligence models about the planet.

OpenAI’s ultimate goal is to build a general artificial intelligence model, capable of reproducing all human capabilities.

At the end of last month, Google launched a video-generating artificial intelligence model called Lumiére. The presentation material shows a sequence of surreal images in surprising quality compared to those generated by other models currently available on the market.

Google’s AI also has no date yet to reach the public.

According to the article published by Google engineers, the images generated by the new technology maintain coherence over time, unlike other platforms available on the market, such as Runway, already tested by Folha, for up to five seconds (80 frames at a frequency of 16 frames per second).

In the case of Google, the strategy was to train the neural network based on a temporal vector, in addition to the composition and colors of the frame. Thus, the time factor comes into play. See how it works in the diagram below.