Meta currently offers an AI video generation service on Twitter called Make-A-Video. While it looks pretty bad right now, the number of comments in just one day suggests that soon the AI image generation craze will be replaced by AI video generation. This is a huge leap forward, and researchers are pushing the limits of generative art as we know it, especially the limits of the amount of data needed to bring an image to life.

"With just a few words, this cutting-edge AI system generates high-quality video from text prompts," Meta AI writes in a tweet, soliciting prompts. Tips to prevent unregulated gore and porn from being generated in large quantities and posted on Twitter" Send a prompt and they may post the results.

Instead of waiting for the Meta AI team to pick your prompt out of thousands of comments (perhaps scarred for life), you can go to the Make-A-Video studio (opens in new tab) and use the Google form to register your interest in this tool (opens in new tab) (opens in new tab) You can also register.

The accompanying research paper (PDF warning (opens in new tab)) calls the Make-A-Video process "an effective method to extend the diffusion-based T2I model to T2V through a spatiotemporal factorization diffusion model." This is a fancy way of saying that they used an evolved version of the diffusion Text-to-Image generative model to move the picture.

"While progress in T2I generation has been impressive, progress in T2V generation has lagged for two main reasons.

Essentially, the size and accuracy of the data sets required to train current text-to-video AI models are too vast to be viable.

The paper notes that the surprising aspect of this evolution is that it "does not require text-to-video paired data. This differs from many video and image generators out there, which rely on galleries of content already paired with text. This is a significant advantage over prior research."

There are several ways to use this tool, either to fill in movement between two images, simply add movement to a single image, or create a new variation of a video based on the original. The results are fascinating. Dreamlike and psychedelic, they can be generated in several different styles.

Admittedly, these are a bit spooky, especially when one recalls that the results are only getting more realistic, but a little hike through the uncanny valley never hurts for Halloween.