FLEXIBLE VIDEO DIFFUSION

Summary from the TLDR AI newsletter

Flex-Forcing trains video diffusion models to switch between bidirectional and autoregressive generation through flexible temporal and denoising-step chunking. The approach improves inference speed, video quality, and long-range stability across different compute budgets.