Woofun AI reports that ByteDance launched Seedance 2.0 at the end of 2025, transitioning its architecture from UNet to DiT to leverage computational scalability. Project lead Zeng Yan directed the training of a 200B parameter model despite internal technical disagreements.

ByteDance deployed an internal evaluation team exceeding 1,000 members, assigning over ten data personnel per algorithm engineer to ensure model efficacy without data distillation. This strategy enabled Seedance to generate more than half of Volcano Engine's supermodel revenue.

Concurrently, a near 90% reduction in AI drama production costs prompted short film companies to purchase full Seedance licenses and reinvest savings into TikTok traffic, completing a commercial cycle of model sales, content production, and consumption.