PACR-Video: Parameter-Efficient Multi-Shot Long Video Extrapolation

Long video generation faces challenges in cross-shot consistency and computational cost. Existing methods rely on full model fine-tuning or memory/streaming mechanisms, often struggling to balance parameter efficiency and long-range coherence. PACR-Video proposes a novel paradigm: freeze a pretrained text-to-video diffusion Transformer, insert low-rank temporal adapters at key positions, and condition them via learned shot-role prompt tokens. These tokens accumulate in a recursive prompt bank storing entity, location, action, and style information. When generating the next shot, the system routes appropriate prompt combinations through adapter gates based on predicted narrative dependencies. Training uses a dual objective: local shot reconstruction and global story-level identity contrast, with prompt sparsity regularization. On six standard benchmarks, PACR-Video significantly outperforms T2V baselines, fine-tuning baselines, memory-augmented, streaming, and recursive-context methods on FVD, CLIPSIM, identity preservation, temporal smoothness, and other metrics. Human preference evaluations also confirm its generated videos are more coherent and natural. This work shows that compact prompt routing and lightweight temporal adaptation provide sufficient controllable capacity for stable long video extrapolation, opening a new path for future controllable video generation.