OPENAI presents Sora 2, more coherent, finer in the details

Sora 2 promises a real gap in terms of realism, complexity of the scenes and introduces sound system. The model is gradually deployed by invitation.

Openai takes a step forward in the video generation. In order to replace its first generation model, already aging against Veo 3, Openai launched this September 30 a new version of Sora. The Californian company evokes an update comparing the transition between Sora to Sora 2 to that of GPT-1 to GPT-3.5. More realistic videos, better adhesion to prompt, sound system … Sora 2 seems to compete with the main actors of the emerging market of video generation.

Better understanding of the physical world

Since the release of Sora last February, the team in charge of Sora at Openai has set out to train the model to simulate the physical world as faithfully as possible. The objective is to make Sora capable of understanding how forces and movements act in the real world to try to reproduce them as finely as possible in generated videos. Gigantic datasets of content (mainly videos) from the web, third -party partnerships or generated by Sora users or researchers have been set up.

According to Openai, Sora 2 is now able, like Veo 3, to generate physically credible videos. Openai even talks about scenes that have so far been impossible to reproduce with a video generation model. The company cites in particular the case of Olympic gymnastics routines, backflips or even complex skating movements (Axel).

Sora 2 also makes it possible to model situations of failure. For example, if a basketball player misses a shot, the ball will bounce on the rear panel rather than teleport to the basket. The model would also Perform in terms of prompt adhesion, that is to say with results more in line with the initial prompt. Finally, persistence would also have been significantly improved to make it possible to produce more coherent scenes over time.

Upload of real world images and videos, sound system

The main novelty, Sora can now inject real world content (Cameos). Whether it is images (already possible with Sora 1) or videos, the model reproduces the physical appearance of the human, the object or the animal sent with good precision, according to Openai. Realistic and without prior training of the model in short. Finally the other novelty concerns sound system. Like Veo 3, Sora 2 is now able to generate synchronized sound effects and dialogues, with a good level of realism. With Cameos mode, the voice of a person recorded on video can also be simulated.

Example: a Deepfake from Sam Altman flying GPUs in a Target store.

Sora 2 is launched for the time being in Research release in North America, by invitation. Members already having access to Sora 2 can share an activation code to authorize new users from Sora.com. Sora 2 is available free of charge with generous limits. Chatgpt pro users will soon be able to use a more advanced version of Sora 2 to generate longer and better quality videos. The video generation API will then be launched in a second step. Openai, however, does not yet communicate a price.

A mobile application for iOS

In parallel Openai launches for consumer users an Sora application designed as a social network. Users can create new video generations and remix the creations of others. The new generated videos can then be shared with the community. Each user has a personalized news wire according to their preferences. The application is launched on iOS, initially in the United States and Canada.

If Sora 2 impresses on paper, Openai keeps the vagueness on the essentials: no benchmark has been communicated to objectively compare the performance of the model in the face of competition. It is also difficult to identify the product strategy of the Californian Scale-Up. Is Sora 2 only targets the general public, or does OPENAI ultimately target the B2B market for content creation and cinema? For the time being, the management remains not very readable.