Make-A-Video PyTorch is an open-source implementation of Meta AI’s Make-A-Video text-to-video generation research in PyTorch, with just under 2,000 stars on GitHub, from prolific independent researcher lucidrains.
When Meta published Make-A-Video — a system that generates video from text by extending text-to-image diffusion with temporal structure — the code was not released. This repository is the community’s answer: a from-the-paper reconstruction in PyTorch, applying lucidrains’ well-known pattern of implementing influential papers faithfully enough that other researchers can train, probe and extend them. The work involves the genuinely hard parts of video generation: attention mechanisms that operate across frames, axial convolutions and the other spatiotemporal machinery needed to keep generated clips coherent rather than a slideshow of unrelated images. The repository was last updated in 2024 and should be read as a research artefact — an educational, hackable implementation of a landmark method — not a product for generating clips; today’s practical text-to-video systems are hosted services or far larger open models. As a way to learn how those systems are put together, though, a readable single-author implementation has real value.
Make-A-Video PyTorch is free and open source under the MIT license. Training or sampling video models is GPU-hungry, and that compute is yours to supply; there is no hosted version or paid tier.
Machine-learning researchers and students who want to understand text-to-video diffusion from the inside — by reading and running a faithful implementation of a foundational paper — rather than simply calling a video-generation API.