Sora: OpenAI's video AI is here
OpenAI's video generator "Sora" is now on the market in North America. Sora creates amazingly realistic-looking videos.
Walk through Tokyo (sample image from the Sora pre-announcement in February)
(Image: OpenAI)
A new generative artificial intelligence (AI) from OpenAI creates realistic-looking videos without sound. The service is called Sora and creates videos up to 20 seconds long on command. Sora is not yet available in German-speaking countries. However, in North America, existing ChatGPT subscribers can enter instructions purely as text or as text enriched with still images or videos. The videos have a resolution of up to 1080p and can be ordered in square, portrait or landscape format.
With Sora, OpenAI is also releasing a smaller sister model called Sora Turbo. It calculates faster, but doesn't produce videos quite as good. ChatGPT Plus subscribers can have 50 low-resolution videos (480p) or fewer medium-resolution videos (720p) generated each month. Subscribers to ChatGPT Pro, which costs 200 dollars, can order ten times as many videos, in higher resolution and of longer duration. At the beginning of next year, OpenAI plans to launch additional tariff models on the market to appeal to different user groups.
Videos by heise
The videos produced fluctuate between frighteningly impressive realism and obvious flaws. OpenAI openly admits that Sora is not perfect: "The Sora version we released has many limitations. It often generates unrealistic physics and struggles with complex processes over long runtimes," the company writes in its announcement. All videos are watermarked by the C2PA (Coalition for Content Provenance and Authenticity), and an OpenAI animation can always be seen at the bottom right.
There are rules
It is not permitted to upload images of minors. Images of other people may only be used as part of the prompt if the people depicted have given their consent. Users must also ensure that they waive all necessary intellectual property rights. Depictions of violence and explicit themes, i.e. adult material, are prohibited. Anyone who uses Sora contrary to the regulations can be blocked temporarily or permanently without receiving a refund of fees already paid, warns OpenAI.
The user interface on the Sora website shows videos that other users have ordered. The exact prompts that led to the respective video can be called up. In this way, OpenAI gives its users an insight into the art of AI prompting.
The Sora videos can not only be ordered with individual prompts; the interface also allows users to put together a storyboard. A series of instructions can be entered to calculate a continuous video with different scenes.
Mischief happens
heise online has not yet been able to try out Sora, but the US-American Youtuber Marques Brownlee was allowed to use it for a week. In his examples, it is not always clear whether a scene was filmed for real or generated by Sora. However, Brownlee also shows examples where mistakes cannot be overlooked.
Text overlays are sometimes a jumble of characters, animals or objects sometimes disappear suddenly and unmotivated. Animals or objects also like to move around in a mess. Sora also struggles with the depiction of the legs of moving animals and physical conditions in general. These shortcomings are less noticeable in abstract representations and cartoon-style films than in videos that aim to be realistic.
Empfohlener redaktioneller Inhalt
Mit Ihrer Zustimmung wird hier ein externes YouTube-Video (Google Ireland Limited) geladen.
Ich bin damit einverstanden, dass mir externe Inhalte angezeigt werden. Damit können personenbezogene Daten an Drittplattformen (Google Ireland Limited) übermittelt werden. Mehr dazu in unserer Datenschutzerklärung.
We know from Brownlee's video that OpenAI also offers downloads without watermarks on request. This probably only refers to the visible animation, but we have not yet been able to determine this. The release of Sora is part of OpenAI's "Shipmas", twelve days of new launches.
(ds)
