Long Video Generation
Overview
Long video generation studies how to produce sequences that extend beyond short clips without losing visual quality, subject identity, scene structure, or narrative coherence. Common approaches extend diffusion models with overlapping windows or autoregressive chunks, while memory retrieval, compressed context, visual anchors, and agent-based planning help carry information across shots. Work also targets bounded memory and faster inference through efficient attention, token compression, and adaptive computation, with applications including storytelling, talking heads, and synchronized audio-video. Minute-scale generation is increasingly explored, but preventing drift while preserving motion diversity and controllability remains difficult, and evaluation still struggles to measure long-range consistency reliably.
Latest papers
67 papers8 papers in the last four weeks, up 33% on the four weeks before. 0.0% of all new papers.