MovieChat

Why we included this project

MovieChat is a CVPR 2024 research system that attacks the context-length problem in video understanding directly. It's a multimodal large language model that compresses dense visual tokens into a sparse memory representation, which is how it reasons over videos with tens of thousands of frames on a single 24GB GPU. The repo ships the model code, training and evaluation scripts, and the MovieChat-1K benchmark with a public leaderboard, so it doubles as a reference point for comparing newer long-video methods. Teams building video Q&A or summarization tools can borrow the memory-compression ideas and the evaluation harness even if they plan to build on a newer base model. It's a research artifact rather than a turnkey service, so expect to adapt the code before it's production-ready.