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

2026-07-03

AI Summary

This work addresses a fundamental trade-off in current video generation models: bidirectional diffusion models ensure strong global consistency but suffer from high computational costs, while autoregressive models are efficient yet struggle with long-range coherence and exposure bias. To reconcile these limitations, the authors propose a unified training and inference framework that introduces a flexible spatiotemporal blocking mechanism, jointly partitioning the time axis and denoising steps within a diffusion model. This enables, for the first time, a single model to support both bidirectional global planning and intra-block autoregressive generation, with the added capability of arbitrary-order, non-causal autoregressive sampling. By relaxing strict causal constraints, the method allows dynamic adaptation of inference strategies based on computational budgets, achieving significant improvements over strong fixed-schedule baselines across multiple video generation benchmarks in terms of generation quality, long-video stability, and inference efficiency.

Abstract

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.

Problem

• video generation
• bidirectional diffusion
• autoregressive modeling
• inference paradigm
• long-range consistency

Innovation

• Flex-Forcing
• video diffusion model
• bidirectional generation
• autoregressive generation
• flexible chunking