Multi-Grid Post-Training for Long-Form Multi-Shot Video Generation
Summary
MovieGrid is a multi-grid post-training paradigm that decomposes long videos into spatially arranged chunks to improve multi-shot coherence and efficiency, achieving state-of-the-art intra-shot and inter-shot consistency in video generation.
View Cached Full Text
Cached at: 09/09/26, 04:29 AM
Paper page - Multi-Grid Post-Training for Long-Form Multi-Shot Video Generation
Source: https://huggingface.co/papers/2609.06373 Authors:
,
,
,
,
,
,
,
,
,
Abstract
MovieGrid decomposes long videos into spatially arranged chunks for joint modeling, improving multi-shot coherence and scaling video length efficiently.
Generating long-form multi-shot videos requires coherent within-shot motion and visually consistent narratives across shots. Existing video generators favor continuous motion and struggle to present complete shot sets when an entire narrative is packed along one temporal axis. We proposeMovieGrid, aMulti-Grid Post-Trainingparadigm that decomposes a long video into shorter, temporally ordered chunks and arranges them on a spatial grid for joint modeling. This design reduces the number of shots handled by each temporal axis while enabling global information exchange across chunks. We construct the Multi-Grid Long Video (MGLV) dataset from 1,000 long-form videos using source video collection, hierarchical segmentation,grid videoconstruction, and character-aware story annotation, producing 54Kgrid videos paired with story prompts. OurNoise-Free Random-Grid Trainingretains a random subset of chunks as clean visual context for denoising the remaining chunks.Grid Embeddingencodes grid structure, character-aware Story Prompts link recurring entities, andGrid Boundary Lossstabilizes layouts. Under the same token budget,MovieGridgenerates 6.05 times more shots than Temporal Packing in a 1,616-frame video. On a benchmark spanning five real-world categories, it achieves state-of-the-artintra-shot consistency(0.9131 versus 0.8086 for HoloCine) andinter-shot consistency(0.5914 versus 0.5384 for StoryMem).MovieGridcan further scale video length with minimal compromise through single or multiple generations.
View arXiv pageView PDFProject pageGitHub0Add to collection
Get this paper in your agent:
hf papers read 2609\.06373
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2609.06373 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2609.06373 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2609.06373 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
GridProbe: Posterior-Probing for Adaptive Test-Time Compute in Long-Video VLMs
GridProbe is a training-free inference paradigm for Long-Video VLMs that adaptively selects relevant frames using posterior probing, achieving sub-quadratic attention costs with minimal accuracy loss.
MilliVid: Hierarchical Latents for Long-Range Consistency in Video Generation
The paper introduces MilliVid, a method for improving long-range consistency in video generation by using a multi-scale autoencoder to compress frames into hierarchical tokens and then generating them with a coarse-to-fine diffusion model, outperforming baselines on Minecraft videos.
Enhancing Train-Free Infinite-Frame Generation for Consistent Long Videos
MIGA is a train-free method for generating consistent long videos by reducing the training-inference gap and enhancing temporal consistency through dual consistency mechanisms.
MVTrack4Gen: Multi-View Point Tracking as Geometric Supervision for 4D Video Generation
MVTrack4Gen introduces a training framework that uses multi-view point tracking as geometric supervision to enhance motion-aware diffusion models, achieving state-of-the-art geometric consistency and motion fidelity in novel-view video generation from monocular video.
MV-Forcing: Long Multi-View Video Generation via 4D-Grounded Spatio-Temporal Self-Forcing
MV-Forcing introduces a diffusion framework that combines temporal and view-wise autoregression to generate long, multi-view consistent videos of dynamic scenes, using a 4D geometric bridge and spatio-temporal distillation to enable arbitrary-length generation from a few-step student model.