OpenAI is strengthening its presence in AI-generated video with Sora, a tool capable of turning text descriptions into realistic videos. Until now less prominent than ChatGPT or DALL·E in this field, the start-up now offers a solution for creating animations up to one minute long, at resolutions of up to 1080p.
Sora is based on conditional diffusion models trained jointly on images and videos of different lengths and resolutions. The technology uses transformers similar to those in large language models to process the spatial and temporal elements of video sequences. The result: the tool can generate complex scenes with several characters, specific movements and detailed backgrounds, while respecting the physical logic of objects.
Despite these feats, OpenAI acknowledges certain limits. Sora can confuse spatial positions (left/right), struggle to track actions over time, or fail to reproduce the properties of complex scenes perfectly. Access to the model is still restricted for now: it is being tested by OpenAI’s red team, as well as by artists, designers and filmmakers to refine its potential. With Sora, OpenAI aims to establish itself in an already competitive market that includes players such as Meta, Microsoft and Runway.
Source: Le Monde Informatique
