Multi-Grid Post-Training for Long-Form Multi-Shot Video Generation

Multi-Grid Post-Training for Long-Form Multi-Shot Video Generation. By Jiawei Mao and ten others, from UC Santa Cruz, the University of Florida, and two other institutions. Published on the sixth of September. Its code repository is worked out: about five hundred source files, with a readme.

A long-form video told in multiple shots needs two kinds of coherence at once: motion that holds together within each shot, and a narrative that stays visually consistent from one shot to the next. Existing video generators favour continuous motion. When a whole narrative is packed along a single temporal axis, they struggle to present the complete set of shots the story calls for.

MovieGrid is a Multi-Grid Post-Training paradigm. It decomposes a long video into shorter chunks, kept in temporal order, and arranges those chunks on a spatial grid so they can be modelled jointly. The layout does two things. Each temporal axis now has fewer shots to handle, and because all the chunks sit together on the grid, information can be exchanged globally across them.

Training data comes from a new dataset, Multi-Grid Long Video, or MGLV. It is built from one thousand long-form videos through source video collection, hierarchical segmentation, grid video construction, and story annotation that is aware of characters. The result is fifty-four thousand grid videos, each paired with a story prompt.

Several components make the grid trainable. Noise-Free Random-Grid Training keeps a random subset of chunks clean, as visual context, while the remaining chunks are denoised. Grid Embedding encodes the structure of the grid. Character-aware Story Prompts link entities that recur across shots. And a Grid Boundary Loss stabilises the layouts.

Under the same token budget, MovieGrid generates six point zero five times more shots than Temporal Packing in a video of one thousand six hundred and sixteen frames. On a benchmark spanning five real-world categories, it reaches state-of-the-art intra-shot consistency, zero point nine one three one against zero point eight zero eight six for HoloCine.

Inter-shot consistency is also state of the art: zero point five nine one four against zero point five three eight four for StoryMem. And MovieGrid can extend video length further with minimal compromise, through a single generation or through multiple ones.

The move is to trade time for space. A generator that handles one long temporal axis poorly handles several short ones laid side by side much better, and seeing them all at once is what keeps characters and look consistent across the story.