This article is also available in:
Français
Once generative AI was good enough for still pictures, generating videos was the next step. Some tools already exist, but they are usually far from perfect: most of the time the camera is almost still, only short clips can be created, and weird artifacts can be spotted. Runway’s Gen-2 is a good example of this.
OpenAI is now tackling this market with Sora, a generative AI that has just been officially unveiled. It can create sequences of up to 1 minute from a text prompt or by relying on an image. Sora can also extend an existing video or fill in gaps of a video where some parts are unavailable.
The results are impressive. Judge for yourself:
More videos are available on the official Sora website.
Impressive, but far from perfect (yet)
Of course, videos generated by Sora are not perfect: upon close inspection, you can notice scale issues, as well as weird artifacts. For example, look closely at the legs of people in the background of the video of the young woman walking in Tokyo, and you’ll lotice these legs sometimes behave erratically. On a video showing an octopus, suckers are placed on the wrong side, arms merge during the video.
Other examples include weird perspective issues (parts of buildings that keep facing the camera as it rotates around said buildings), a train whose carriages have different lengths (just count the number of windows), a chair changing shape, a crab with weird legs…
Despite issues, Sora is impressive
Sora does not provide perfect results, but the videos it can generate are still especially realistic. In many cases, they could deceive the viewer on the first watch. Which, for many use cases, is sufficient.
Sora: what about ethics?
Obviously, Sora raises a lot of questions, especially in a context where AI, deepfakes are already regularly used to create misinformation or to target personalities.
OpenAI tries to be reassuring, and explains that the company is “working with red teamers — domain experts in areas like misinformation, hateful content, and bias — who will be adversarially testing the model”. In other words, the goal will be to try to generate unethical content, in order to stop Sora from generating it.
Furthermore, OpenAI explains that it is building tools that will, for example, allow detecting whether a video was created using Sora. They also “plan to include C2PA metadata” should the model be used in an OpenAI product.
OpenAI adds that existing text classifiers will be used: the idea here being to analyze and reject a prompt that would be detected as violating usage policies. This, in theory, should help filter out requests for content such as “extreme violence, sexual content, hateful imagery, celebrity likeness, or the IP of others”.
Last, but not least, OpenAI explains that tools will review the frames of each video generated by Sora before showing them to users.
Of course, should Sora becoma a commercial product, we’ll have to check wether all these promises are kept.
We should also highlight that on the official Sora website, OpenAI does not give much information about the training data used by the company. Yet, this is a crucial point (as recent news has shown). This choice contrasts with other companies like NVIDIA, Adobe, which state that they only used data gathered in a ethical manner.
Another point not mentioned by OpenAI is the ecological impact of Sora, as training such models requires a lot of resources.
What’s next?
For now, Sora remains a closed tool and not a commercial product. You cannot use it. However, a technical report is available. It include more information about the technology used to create Sora, as well as other examples of videos, and references to scientific papers.
In any case, Sora reignites the debate on the use of generative AIs. Given the spectacular progress made in recent months, it is reasonable to think that the flaws we mentionned earlier will become less visible over time.
This opens up the possibility of creating movies, series, TV shows, YouTube videos without a camera or 2D/3D animation software. With all the advantages that this entails… But also massive ethical risks.
More than ever, legal, ethical issues, and reflections on the impact generative AI will have on the creative industry, must be at the heart of considerations. Advancing these reflections, as well as new regulations, as quickly as the progress of generative AIs, is a challenge in itself.
We’ll keep you updated on this matter in the weeks and months to come.
