It has been reported that OpenAI is developing a very new edition of its video-generating AI, Sora, with an entire upgrade. They are saying that they’ve began training from scratch to enhance performance in comparison with the version released last February, and because of this its release has been delayed constantly.
The Information reported on the twenty fifth (local time), citing sources who spoke with OpenAI’s executives, that Sora is within the technique of developing a new edition that may train with a brand new dataset consisting of hundreds of thousands of hours of video.
Accordingly, OpenAI expects the brand new Sora to not only have the perfect quality, but in addition solve most of the prevailing problems.
It focuses on producing higher quality, longer clips faster than the version announced in February. On the time, Sora was said to have the opportunity to supply one-minute videos, which continues to be the longest amongst current video models.
Additionally it is reported that the video data used for Sora’s training was meticulously chosen. Which means high-resolution videos with various styles and topics were collected. This was explained as a measure to not only increase Sora’s output resolution, but in addition to avoid bias toward certain styles.
The evaluation is that the rationale OpenAI didn’t release Sora, which was introduced in February, was since it was not at a level suitable for business release. For instance, the previous version took greater than 10 minutes to generate a 1-minute video.
One other creator, Sedeberg, who used Sora, said that it took lots of of attempts to get the perfect image as a result of the uneven output quality, and that it was difficult to take care of object consistency. There have been also cases where the laws of physics were ignored, resembling chairs suddenly appearing, and cases where the human body was not properly represented, resembling the variety of fingers.
To unravel all these problems, OpenAI decided to rebuild the dataset from scratch and train Sora anew. For this reason it has not been released to the general public even after six months.
Sora is not the primary OpenAI release to be delayed. GPT-4o’s Advanced Voice Mode (AVM), which was officially released this week, also took 4 months to be released after its public release in May.
Specifically, it’s unclear when Sora, which has began training again with a brand new dataset, might be released. Nevertheless, if development is accomplished as OpenAI intends, the ripple effect is predicted to be even stronger.
Reporter Im Dae-jun ydj@aitimes.com
