NVIDIA's research team has released Flex-Forcing, a method that lets a single video diffusion model switch generation modes at inference time. It unifies two approaches, bidirectional diffusion and autoregressive generation, in one model that can be selected according to the compute budget, and has been accepted as a Spotlight at the machine learning conference ICML 2026.
Continue reading
The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in✓ Signed in — this article isn’t included in your current plan.