Act2Cut: Continuous Multi-Shot Video Narrative Match on Action-Cut
Abstract
Shot matching cut is pivotal for cinematic inter-shot coherence. Current multi-shot video generation primarily focus on superficial visual consistency while often neglecting the fundamental action continuity essential for professional filmmaking, a technique known as Match on Action (Action-Cut). It requires maintaining both action state consistency and kinetic momentum continuity between adjacent shots. Consequently, state-of-the-art models frequently exhibit severe artifacts during transitions, including action resetting, temporal flickering, and physical logic disruptions. To address these limitations, we propose Act2Cut, the first unified framework specifically designed for continuous-action-driven multi-shot video generation. Architecturally, we introduce Global-Local-Shot Rotary Position Embedding (GLoS-RoPE), a hierarchical temporal RoPE that preserves global narrative flow while managing local shot evolution. To anchor physical integrity, we propose Transitional Boundary Residual (TBR) and a corresponding boundary loss to enforce action state consistency between adjacent shots’ tail and head frames. Furthermore, Shot Causal Mask (SCM) and Hierarchical Context Mask (HCM) are integrated to model inter-shot causality and eliminate semantic leakage across distinct cinematic descriptions. These components decouple global action semantics from local shot cinematography by intervening in the DiT attention layers. For training, we construct ActCutVid-1M, a large-scale, hierarchically annotated dataset featuring decoupled cinematographic languages and Focus-of-Attention (FoA) guidance, sourced from films, TV series, and AI-generated content. Based on this, we establish ActCutBench, a specialized benchmark for Action-Cut task, which evaluates narrative consistency, visual coherence, action continuity, and cinematographic language alignment. To unify diverse generation paradigms within one model, we employ multi-task tuning to enable native one-pass generation and prediction for Text-to-Video, Image-to-Video, Storyboard-to-Video, and Video-to-Video. Extensive experiments demonstrate that Act2Cut effectively maintains consistency and coherence while enabling action continuity.