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

Recent progress in large-scale generative models has substantially advanced video generation, yet existing methods remain constrained by a rigid inference paradigm. Bidirectional diffusion models excel at global coherence and visual fidelity but suffer from slow inference, while autoregressive models offer efficient and streaming generation at the cost of long-range consistency and exposure bias. We introduce Flex-Forcing, a unified training and inference framework that enables a video diffusion model to seamlessly operate under both bidirectional and autoregressive generation regimes. The core idea is a flexible chunking mechanism jointly defined over the temporal axis and denoising steps. This design allows the model to (1) perform flexible chunking according to different device budgets, (2) perform bidirectional inference across chunks for global structure planning, while generating frames autoregressively within each chunk for efficient and fine-grained synthesis, and (3) perform any-order, any-timestep autoregressive generation without the strict causal constraint. Extensive experiments on multiple video generation benchmarks demonstrate that Flex-Forcing achieves consistently better video quality, long-video stability than strong baselines with a rigid inference schedule, while offering faster inference.

Video diffusion models are restricted by a rigid tradeoff. Bidirectional models employ full-context self-attention for high visual coherence but suffer from severe computational overhead. Autoregressive models utilize causal masking and Key-Value caching for high-speed streaming generation but accumulate compounding approximation errors, causing structural drift and exposure bias over time. Flex-Forcing unifies both paradigms via Flexible Chunking across two orthogonal dimensions: temporal frames and denoising timesteps. It divides videos into dynamic partitions, applying bidirectional attention inside each chunk for structural planning, while generating between chunks autoregressively for speed. A timestep-conditioned K-Projection aligns clean cached key states into the noisy latent space to resolve signal-to-noise ratio discrepancies. Flex-Forcing expands the efficiency-quality Pareto frontier, outperforming rigid baselines like Self-Forcing. It dynamically adapts to hardware constraints and unlocks non-causal downstream utility, enabling localized, any-order video editing without breaking global temporal consistency.