Seedance

In February 2026, a video of Tom Cruise and Brad Pitt fighting — never filmed, never staged — racked up millions of views on social media within 48 hours. It was generated by a single model from a single text prompt. That model was Seedance 2.0, and it had just changed the conversation about AI video forever.

Origin: The Seed That Grew Into a Studio

Seedance is the video generation flagship of ByteDance's internal research division, known simply as Seed. Established in early 2023 in the wake of ChatGPT's global disruption, Seed was ByteDance's strategic bet on foundation model leadership — not just for language, but for every modality. The parent company, best known for TikTok and Douyin, had a singular advantage: it operated the world's most-consumed short-form video platform and had trained its recommendation algorithms on hundreds of billions of viewing decisions. That understanding of what makes video compelling at human scale was embedded from the start into Seedance's training and evaluation philosophy.

Architecture: Dual-Branch, Native Audio-Visual

Seedance 1.0, published in June 2025 by a 44-researcher team, introduced a Diffusion Transformer backbone with multi-source data curation, native multi-shot generation, and a 10× inference acceleration achieved through multi-stage distillation. It immediately topped the Artificial Analysis AI Video Leaderboard in both text-to-video and image-to-video categories. Seedance 1.5 Pro, released December 2025, became the first commercially available model to generate video and audio natively in a single unified pass — eliminating the post-production stitching that had plagued all prior AI video tools. Seedance 2.0, launched February 12, 2026, introduced a Dual-Branch Diffusion Transformer architecture that accepts up to 12 simultaneous reference files across four modalities: text, image, video, and audio. It generates clips up to 20 seconds at 2K resolution (2048×1080) with phoneme-level lip-sync across 8+ languages, director-level camera control, and physics-aware motion that handles collisions, fabric dynamics, and fluid behavior with unprecedented realism.

The Controversy That Defined Its Debut

Within 48 hours of launch, Seedance 2.0 generated a wave of viral deepfakes — celebrity likenesses, copyrighted franchises, beloved film characters — leading Disney, Paramount, the Motion Picture Association, and SAG-AFTRA to issue cease-and-desist letters. ByteDance paused the planned global API rollout on February 24, 2026, to implement stricter IP safeguards. As of March 2026, the model remains available within China via Jimeng (即梦) and Doubao, with international access restricted. The controversy has done little to diminish its technical reputation: industry analysts have called it "the second DeepSeek shock," noting its ability to match Western competitors at a fraction of the compute cost.