NVIDIA Proposes Flex-Forcing for Unified Video Generation

nvidia · hf · 2026-07-08

NVIDIA researchers proposed the Flex-Forcing method, which flexibly chunks timesteps and denoising steps. This allows a single video diffusion model to operate in both bidirectional and autoregressive generation modes without separate training. Experiments show that this method improves video quality while accelerating inference speed, marking a significant step toward a unified video generation framework.

Related event: NVIDIA Introduces Flex-Forcing for Video Generation(2 posts)→

Original post →

More from Multimodal

Multimodal channel →