Report: ByteDance Is Building a Real-Time Spatial Video Model on Seedance, Targeting 20 fps

ByteDance is reportedly readying a real-time spatial video model built on Seedance, according to a Bloomberg report the company has not confirmed.

For creators, the difference from today's Seedance would be fundamental. Seedance 2.5, which ByteDance's Seed team launched on July 31, 2026, produces finished clips of up to 30 seconds per generation. The reported model would instead render a scene continuously while a viewer moves through it, which is closer to a game engine than to a clip generator.

Bloomberg's Haze Fan reported on September 7 that founder Zhang Yiming is personally overseeing the project, citing people familiar with the matter who asked not to be identified. Bloomberg's site is closed to automated readers, so this article draws on the Bloomberg story as republished by The Star and on accounts of it from The Next Web and TechNode.

According to the report, the Seedance-based model would let users create interactive virtual worlds for livestreams, short-form dramas and games. It is reportedly meant to generate worlds that respond to the voices and movements of people wearing headsets from Pico, ByteDance's extended reality arm. One of Bloomberg's sources put the target at on-demand video of 20 frames per second with latency of about 0.05 seconds, or 50 milliseconds.

The report adds that generation would happen in the cloud rather than on the headset. That would cut the processing power a headset needs and could open the way to cheaper hardware. Bloomberg likened the experience to Google's Genie, which it said launched earlier this year and lets users interact with a world rendered in real time.

A launch could come as soon as the month after the report, meaning October 2026. One of Bloomberg's sources cautioned that the timing is not certain and the plans may change.

The direction would not come from nowhere. TNW, relaying earlier reporting by Chinese outlet 36Kr, wrote that ByteDance made world models its top AI priority for 2026 and aimed to ship at least one by the end of the year. On the record, ByteDance's own Seedance 2.5 post says the model is already used to generate synthetic video for training robots and for simulating rare conditions in autonomous driving. That post says nothing about real-time or spatial output.

A ByteDance spokesperson did not respond to Bloomberg's request for comment. TNW stated that it has not independently verified the account, and TechNode's short summary credits Bloomberg alone.

Still, everything here rests on a single report built on anonymous sources. Neither TNW nor TechNode adds independent confirmation. ByteDance has not announced a product name, price, demo footage or release date, and the newest post on the Seed team's blog as of September 19 is dated August 5 and concerns a different model. The 20 fps and 50-millisecond figures are reported targets from one person, not published specifications.

What creators can use today is Seedance 2.5, which ByteDance says is rolling out on Jimeng AI and Doubao Pro with API access through BytePlus ModelArk to follow; Runway's API changelog shows Seedance 2.5 available to developers there since August 7, 2026.