Sora (text-to-video model) facts for kids
Developer(s) | OpenAI |
---|---|
Initial release | 15 February 2024 |
Platform | OpenAI |
Type | Text-to-video model |
Sora is a special computer program made by OpenAI. OpenAI is a research group in the U.S. that works on artificial intelligence (AI). Sora can create videos from text descriptions you give it. It can also make existing videos longer, either forwards or backwards in time. This type of AI is called a text-to-video model.
Contents
How Sora Was Developed
Before Sora, other computer programs could also make videos from text. These included Make-A-Video from Meta, Gen-2 from Runway, and Lumiere from Google. Lumiere was still being researched in early 2024.
OpenAI also created a program called DALL-E 3. DALL-E 3 makes images from text. It was released in September 2023.
The team that made Sora named it after the Japanese word for "sky." They chose this name to show its "limitless creative potential."
Sora's First Look
On February 15, 2024, OpenAI first showed off Sora. They released many short videos that Sora had created. These videos were high-quality and looked very real. Some examples included:
- An SUV driving on a mountain road.
- A fluffy monster next to a candle.
- Two people walking in the snow in Tokyo.
- Fake old videos from the California gold rush.
OpenAI said that Sora could make videos up to one minute long. After showing the videos, the company shared a report. This report explained how they taught the computer program to create these videos. Sam Altman, the CEO of OpenAI, also used Sora to make videos for people on Twitter. He would create videos based on their ideas.
OpenAI plans to let everyone use Sora in the future. However, they have not said exactly when this will happen.
Testing Sora Safely
OpenAI gave a small group of people early access to Sora. This group was called a "red team." These experts looked for problems like misinformation or unfairness in the videos Sora made. They helped test the program to make sure it was safe.
The company also shared Sora with artists and video makers. They wanted to get ideas on how Sora could be useful for creative work.
How Sora Works and What It Can Do
The technology behind Sora is similar to what is used in DALL-E 3. Sora uses a method called "denoising diffusion." It creates videos by cleaning up noisy 3D "patches" of information. Then, it turns these patches into a standard video.
OpenAI taught Sora using many videos. Some were publicly available, and others were licensed for this purpose. However, OpenAI did not say exactly how many videos they used or where they all came from.
Sora's Current Limits
When Sora was first released, OpenAI noted some things it still struggles with:
- It has trouble making things move exactly like they would in the real world.
- It sometimes doesn't understand cause and effect. For example, if someone bites a cookie, the cookie might not show a bite mark.
- It can get confused between left and right.
OpenAI also has rules to keep Sora safe. Sora will not create videos that are violent or hateful. It also avoids making videos of real celebrities or content that uses existing copyrighted characters.
Tim Brooks, a researcher who worked on Sora, said that the program learned to create 3D graphics on its own. It did this just from the videos it was shown. Another researcher, Bill Peebles, noted that Sora could automatically create different camera angles without being told to.
To help people know if a video was made by AI, Sora-generated videos are tagged. This tag is called C2PA metadata. It shows that the video was created by an AI program.
See also
- VideoPoet