When throwing Sora 2, Operai Kned Two Strokes with a stone. As you approach its tenth anniversary, the Sam Altman company presents a new important version of its artificial intelligence video generator released in February 2024.
The experience is articulated around an application dedicated to iOS, which has everything from a social network, with the difference that all the content is created precisely only by AI. In addition, an “Cameo” option allows you to integrate into these videos, whatever the style chosen at the base, realistic, cartoons, etc.
A model that best manages physics
Because with the Sora 2 model, Openai claims to have greatly improved the physics of the bodies, objects and elements. It also allows you to better manage the somewhat eccentric indications, and then offer a more realistic video. Similarly, OpenAi seems to have corrected one of the oldest models that are said to be “too optimistic.” That is, they will tend to modify elements, the trajectory of an object, in summary, to distort reality to bend in order. For example, a basketball ball that teleports a few centimeters to enter a basket that was going to be lost. Sora 2 will assure you that the balloon bounces in the basketball panel, for example. But Openai insists that it is important to model both success and the failure of an action to produce a “useful world simulator.”
Another way to achieve a more realistic representation, the ability of the model to follow and administer complex instructions distributed on several levels, so that the action is continuous and consistent. It has been reinforced, and if everything obviously is not yet perfect, progress is obvious.
Compared to the first model, which gave rise to Bizzareries, Sora 2 intends to be “the best video generator in the world.” It is in any case in these terms that this model is described by Gabriel Petersson, a researcher within the Sora Division, Operai.
In X, he shared a short video in which you can see it set up a dragon, escape from the police, immerse yourself in the water. In each plan, his face is recognizable, from beginning to end. The result is impressive, but nevertheless raises the question of the data. If you “give” your face to OpenAi, what will it be? Will it be used then for the training of the models, replicated?
In addition, we believe that improvements remain possible, particularly in the management of small objects or certain unrealistic situations. In the video of Gabriel Petersson, sometimes we see a man taking a wall and “disappear” inside, sometimes an object that does not move when one leg touched it. Small details that still allow it to see that the video is generated by AI, but how long?
Note that to use it, Sora 2 requires a chatgpt plus (22.99 euros per month) or chatgpt pro
Source: BFM TV
