Consistency-Preserving Diverse Video Generation
A new joint-sampling method improves batch diversity by 30% without sacrificing smooth motion or requiring costly decoder backpropagation.
Researchers Xinshuang Liu, Runfa Blark Li, and Truong Nguyen developed a joint-sampling framework for flow-matching video generators. It applies diversity-driven updates, then removes components that would decrease temporal consistency. Crucially, it computes objectives with lightweight latent-space models, avoiding expensive video decoding and decoder backpropagation. Experiments on a state-of-the-art text-to-video model show it matches strong diversity baselines while substantially improving temporal consistency and color naturalness.
Why It Matters
This makes generating multiple, high-quality video variations from a single prompt more efficient and practical for creators.