Make-A-Video
By Meta
Make-A-Video is a text-to-video generation model developed by Meta that extends a pretrained text-to-image diffusion model into the video domain by adding spatiotemporal layers, allowing it to learn motion from unlabeled video data without…
Definition
Make-A-Video is a text-to-video generation model developed by Meta that extends a pretrained text-to-image diffusion model into the video domain by adding spatiotemporal layers, allowing it to learn motion from unlabeled video data without needing paired text-video training examples. It was published as a research system with sample clips rather than released as an open-weight or public product, and it remains a frequently cited reference point for later text-to-video architecture research.
Overview
Make-A-Video is a text-to-video generation model developed by Meta that addresses a practical data bottleneck in video generation: paired text-and-video training data is far scarcer than paired text-and-image data, and gathering enough of it to train a video model from scratch the way text-to-image models are trained is difficult. Make-A-Video sidesteps this by starting from a model that already understands text-to-image generation. The approach extends a pretrained text-to-image diffusion model by adding spatiotemporal layers, network components that let the model reason about motion and consistency across frames, and then trains those added layers using unlabeled video footage, which is comparatively easier to obtain than footage paired with matching text descriptions. The model learns how things move in the world from the unlabeled video, while retaining the pretrained model's learned understanding of how text relates to visual content. This split, learning appearance and text-grounding from paired image-text data and learning motion from unpaired video separately, is the specific mechanism that lets the overall system be trained without the large paired text-video corpora that a fully end-to-end video model would otherwise require. Within text-to-video approaches, this strategy of extending a pretrained image model with added temporal layers is distinct from training a video-native architecture from the outset, as CogVideoX does with a diffusion transformer trained on video directly, and it is conceptually related to Emu Video's later two-stage approach of generating an anchor image before animating it, though Make-A-Video's temporal-layer extension and Emu Video's explicit two-stage pipeline are different specific mechanisms. In practice, Make-A-Video has been presented through Meta's research publications with example short clips demonstrating text-to-video generation, illustrating that reusing an image model's learned text-image associations combined with motion learned from unlabeled video could produce plausible short video clips without needing large paired text-video datasets. Its limitations are typical of early text-to-video work: generated clips are short, complex or physically implausible motion can look unnatural, and because the underlying text understanding comes from the image model rather than from text paired directly with video, association between specific described actions and the resulting motion can be looser than a model trained end-to-end on matched text-video pairs would achieve. As one of the earlier publicly documented text-to-video systems, Make-A-Video is frequently cited in later work as a reference point for the pretrained-image-model-plus-temporal-layers strategy, a lineage that subsequent research, including some approaches that instead train video-native architectures from scratch, has explicitly built on or contrasted itself against.
Key Concepts
- Extends a pretrained text-to-image diffusion model into video generation
- Learns motion from unlabeled video data without paired text-video examples
- Adds spatiotemporal convolution and attention layers to the base architecture
- Includes frame interpolation to smooth generated motion
- Published as a Meta research system rather than an open-weight release
- Demonstrates a data-efficient path to text-to-video training
- Predates and differs architecturally from Meta's later Emu Video
Use Cases
Frequently Asked Questions
From the Blog
Chain-of-Thought Prompting: Make AI Reason Better
Chain-of-thought prompting asks a model to reason step by step before answering, which improves accuracy on problems that need multiple stages of logic.
Read More AI & TechnologyHow to Make LLM-as-a-Judge Scoring Reliable
A judge model is only trustworthy once you have calibrated it against human labels and controlled for its known biases. This covers writing rubrics with observable criteria, swapping positions in pairwise comparisons, measuring agreement with humans, and knowing when not to use a judge at all.
Read More AI & TechnologyMultimodal AI: Vision, Audio, and Beyond
Modern AI models can see, hear, and reason across text, images, audio, and video simultaneously. This guide explains how multimodal AI works, what's possible in 2026, and how to use vision and audio capabilities in real applications.
Read More AI & TechnologyMultimodal AI Explained: Text, Images, and Beyond
Multimodal AI processes and connects several data types like text, images, audio, and video all at once. Here is how it works and why it matters now.
Read More