When ByteDance's video model Seedance 2.0 was greenlit in late 2025, head Zeng Yan pushed through skepticism to train a 200-billion-parameter model. The team was split on the technical direction, but ultimately converged from UNet to a native DiT (Diffusion Transformer) architecture that better leverages compute scaling.
To ensure quality, ByteDance refused data distillation and instead built an in-house evaluation team of over 1,000 people, giving every algorithm engineer more than a dozen data workers. This combination has made Seedance a leader, now contributing over half of Volcengine's large model revenue.
Since AI-produced short dramas cost nearly 90% less than live-action shoots, many short drama studios are buying the full Seedance model from Volcengine and reinvesting the savings into ad spending on Douyin (China's TikTok). ByteDance has thus locked in a unique flywheel: selling AI models, fueling short drama production, and consuming ad inventory.