Flex-Forcing: Towards a Unified Autoregressive and Bidirectional Video Diffusion Model

Flex-Forcing, from NVIDIA Research with Xinchao Wang and Arash Vahdat among the authors, reframes video diffusion as a single model that can switch at test time between bidirectional and autoregressive generation. Built on Wan2.1-T2V-1.3B and trained with a 1.3B teacher, it introduces flexible chunking over both frames and denoising timesteps, so early high-noise steps can use coarse global chunks while later steps use finer local chunks. A key technical addition is K-Projection, a timestep-conditioned key-cache alignment that resolves the signal-to-noise mismatch when causal past tokens and non-causal future tokens coexist in attention. On 5-second video generation, the paper reports that Flex-Forcing reaches 85.07 VBench Total Quality and 86.33 Semantic at 25.8 FPS with 5 denoising steps, while a faster setting reaches 29.4 FPS at 84.63 Total Quality and 85.89 Semantic. In the 2-step regime, it still attains 84.20 Total Quality and 85.27 Semantic at 24.9 FPS. For 30-second video on VBench-Long, it improves total score to 84.01 and raises FPS to 24.96, outperforming Self-Forcing and Infinity-RoPE on long-horizon coherence, motion smoothness, and dynamic degree. The model also enables any-order, any-step autoregressive editing, allowing localized video edits without regenerating the full sequence, which is a practical step toward controllable long-video synthesis.