The Chinese company ByteDance is most well known for the TikTok app. However, the company has now released the latest version of its AI video generator, Seedance 2.0.
Audio-video joint generation
ByteDance says that Seedance 2.0 adopts a unified multimodal audio-video joint generation architecture. In simpler terms, this means that the AI is designed to learn how images and audio work together in the real world. As a result, Seedance 2.0 should create videos which more realistically represent real-world physics. You can use text, image, audio and video prompts for Seedance 2.0 to generate output including video with audio. ByteDance says that Seedance 2.0 features “exceptional motion stability” and delivers “an ultra-realistic immersive experience”.
Sample videos
The sample videos on the Seedance 2.0 website are very realistic. They also feature dramatic and cinematic camera movements. One video opens with a close up of a ballerina’s feet then pulls back to a wide shot as she spins before ending on a mid-shot. Another video begins with an ultra-wide shot of a crowded amphitheatre before moving in like a drone shot to a mid-shot of circus performers. In addition, there is a video which begins with an action shot of a volleyball player hitting the ball. The camera then follows the ball as it flies through the air, transforming into a basketball going through a hoop. Finally, the ball becomes an 8-ball going in the pocket of a pool table, all seamlessly.
Beta release
If you go to the Seedance website, you can only access Seedance 1.5 Pro or Seedance 1.0. That’s because Seedance 2.0 is currently only available to beta users. At the moment, Seedance 2.0 is shown as a “coming soon” on the website. There’s no indication of when the new version will be more widely available.
Additionally, there has been some question about whether or not Seedance 2.0 will be released to U.S. users. This is because ByteDance was forced to give up control of TikTok by the U.S. government due to concerns over the company’s alleged links to the Chinese government.
What we think
The video output from Seedance 2.0 is undeniably impressive. While still having that “uncanny valley” look that affects all AI video, the samples include dramatic camera movements that accurately reflect real-life physics. Seedance 2.0 seems to raise the bar for generative video AI models and will no doubt be very popular with creators who use it. However, some concerns have been expressed over the sources of the data used to train the Seedance 2.0 AI models. Also, some of the sample videos shared online appear to include intellectual property for which permission may not have been given. The legal and ethical debate over generative AI is set to continue for the foreseeable future.
