OpenAI has released its first video creation artificial intelligence (AI) model. In response to the amazing video quality, the CEO of Runway, a leader in this field, commented, “The game is on.” It is evaluated that it is worthy of causing a seismic change in the video creation AI field, where competition is fierce.
Venture Beat reported on the 15th (local time) that OpenAI has unveiled 'Sora', an AI model that generates videos up to one minute in length when text is entered.
“Open AI can quickly create videos up to one minute in length, including multiple characters, specific types of movements, and complex scenes,” he said. “With its deep understanding of language, it can accurately interpret prompts and create attractive characters that express vivid emotions.” “It can be done,” he said.
Currently, most video creation AI outputs last only 15 to 20 seconds. Moreover, the quality of the images created by Sora is amazing.
When the related video was released, Runway CEO Cristobal Valenzuela posted a short message on X (Twitter) titled ‘Game on.’ Runway is a company that is considered a leader in video creation AI.
Cristobal Valenzuela Runway CEO’s post (Photo=X)
Sora, like 'ChatGPT', is based on Transformer architecture. OpenAI explained that Sora was developed based on past research on the image creation AI ‘Dali’ and the GPT model. Sora expresses videos and images as a collection of small data units called ‘patches’, and each patch operates in a similar way to a ‘token’ in a language model.
You can create the entire video at once or expand the created video to make it longer. Provides the ability to predict multiple frames at once. This also solved the tricky problem of keeping the subject the same even if it temporarily disappears from view.
Sora can create videos using only text descriptions or existing still images. It also has the ability to import existing videos and expand them or fill in missing frames.
Of course, like most video AIs, Sora is not perfect yet. OpenAI stated that it may have difficulty depicting complex, physical scenes and that there are cases where it cannot understand cause-and-effect relationships.
For example, following a scene where a person takes a bite of a cookie, the bite marks may not appear. They also reported that they may confuse left and right, or the temporal order of events that occur while following a specific camera trajectory.
Currently, OpenAI is only releasing Sora to a limited number of creative professionals such as visual artists, designers, and filmmakers for the time being and is receiving feedback.
In particular, we plan to evaluate its safety by entrusting it to a 'red team' that looks for flaws and vulnerabilities in the AI system prior to full launch. This is because there is a possibility that video creation services such as Sora can be abused for fake news or deepfakes.
In addition, OpenAI announced that it is also developing a tool to identify Sora-generated videos.
Video created with Sora (Photo = Open AI)
With the advent of Open AI, the competition to develop video creation AI is expected to heat up.
Runway was immediately designated as the industry leader by launching 'Zen 2' with groundbreaking features in June last year. In November, Stability AI unveiled ‘Stable Video Diffusion’ and Meta unveiled the ‘Emu’ video generator. It is reported that Midjourney will also release a video model soon.
Also, last month, Google announced a new space-time diffusion model 'Lumière', MS released a video creation model 'Dragnoir' that manipulates the movement of video frames, and ByteDance released 'Magic Video-V2)'.
0 Comments