Make-A-Video

Make-A-Video is a research system from Meta AI that generates short video clips from text prompts. It builds on Meta's text-to-image work by learning what the world looks like from paired text-image data, and separately learning how the world moves from unlabeled video footage, then combining the two so a written prompt like "a dog wearing a superhero outfit flying through the sky" produces a matching video. It was trained on more than 2.3 billion text-image pairs from the LAION-5B dataset along with video from the WebVid-10M and HD-VILA-100M datasets. Beyond pure text-to-video, the system can animate a static image, interpolate between two images to create a video that links them, and generate variations of an existing video. Output clips are short: 16 frames rendered at 64x64 pixels and upscaled to 768x768, running about five seconds and depicting a single action or scene. Meta reports the system performs better on human evaluations of quality and text-faithfulness than prior text-to-video approaches available at the time of its research paper. Make-A-Video is not a publicly released product. Meta states it is still analyzing, testing, and trialing the system before any public release, and applies watermarking to generated clips to flag them as AI-made. The official site only offers a signup form for future release updates, with no working demo, API, or pricing available.