ByteDance is building AI systems that can generate video in real time with spatial awareness. The TikTok parent company has made world models, which encompass real-time interactive video and 3D-aware generation, a leading AI priority for 2026.
The technical building blocks
ByteDance’s Seed research team has been stacking releases at a steady clip. Seedance 2.0 landed in February 2026 with a focus on physics-accurate multimodal generation. Then came Seedance 2.5 on July 31, which pushed the envelope further with 30-second single-pass audio-video clips, enhanced spatial controls, and a clay-render referencing system for 3D motion planning.
An arXiv paper published around August 24 introduced autoregressive adversarial post-training, or AAPT. The model described in the paper can generate video at 24 frames per second with low latency, producing minute-long coherent outputs.
Before Seedance 2.5, ByteDance had already released Helios, an open-weight model that hit approximately 19.5 fps for minute-long videos running on a single GPU. Helios dropped in March 2026.
Interactive controls are also part of the package. The systems accept pose and camera inputs, enabling applications like virtual human generation where users can direct the output rather than just prompt it.
World models and the Genie 3 benchmark
In June 2026, ByteDance formally elevated world models as a top-tier AI focus. The stated goal is to match Google’s Genie 3, a benchmark for interactive world simulation.
ByteDance is putting serious money behind the effort. Training data budgets for world models have reached eight figures in RMB.
The company’s MultiMedia Lab is simultaneously commercializing 4D and 6DoF (six degrees of freedom) spatial video pipelines for volumetric content. These systems handle both live and on-demand spatial experiences, with deployment progress noted from late 2025 through 2026.