AI image generation has been let loose and it seems there's no going back. With DALL-E 2 now open to all, another player has entered the fray not wanting to lose out – and it's none other than Facebook's parent company Meta. And while DALL-E 2 currently works its magic only with static images, Meta's revealed that it's working on a similar tool for video.
Like with AI image generators such as DALL-E 2, users will be able to type in a descriptive text prompt, and the tool will generate four output options. Named Make-A-Video (give them a break, they were too busy with the tech to work on names) isn't yet public, but Meta AI has been doing requests on Twitter. The results are as creepy as they are astonishing. If you need to catch up on how AI image generation works, see how to use DALL-E 2. We've also seen the best AI art generators compared.
We’re pleased to introduce Make-A-Video, our latest in #GenerativeAI research! With just a few words, this state-of-the-art AI system generates high-quality videos from text prompts.Have an idea you want to see? Reply w/ your prompt using #MetaAI and we’ll share more results. pic.twitter.com/q8zjiwLBjbSeptember 29, 2022
AI art generation is already proving controversial, and that's before it makes the leap to video. We've recently seen the first copyrighted AI art, and we've seen an AI win an art competition. AI-generated images are already all over social media, and we're likely to see even more now that Open AI has opened DALL-E 2 access to everyone.
But what about video? We've had glimpses of AI video generators that other companies are working on, but Meta's looks more advanced than anything we've seen so far, both in resolution and variety. Meta AI says that “in all aspects, spatial and temporal resolution, faithfulness to text, and quality, Make-A-Video sets the new state-of-the-art in text-to-video generation."
But by advanced we don't mean at all normal-looking. As with still AI-generated images, the results are amazing but also a little unnerving, making us wonder if AI-generated art is going to make weird and creepy the new norm. The video looks a little like stop-motion and the glitches make it seem otherworldly and surreal. There's also the fact that they seem to mainly be trying it to dress animals as superheroes or put them in boardrooms.
Every new form or movement in art has its inceptive landmark pieces. Cervantes' Don Quijote is generally considered the first novel. Louis Le Prince's Roundhay Garden Scene is believed to be the first film. Perhaps in years to come people will study Meta AI's 'Confused grizzly bear in calculus class' (above) as a germinal moment inspiring a long line of AI-generated video masterpieces. Meta's taking requests on Twitter for prompts to try with Make-A-Video (see below), and they mainly involve animals too.
#MetaAIMakes pic.twitter.com/xoj93nc0IySeptember 30, 2022
Maybe my favorite AI generated video so far. Prompt: "A fluffy baby sloth with an orange knitted hat trying to figure out a laptop, close up, highly detailed, studio lighting, screen reflecting in its eye.mp4" killer work @MetaAI !! pic.twitter.com/Lvlrl3rWdGSeptember 29, 2022
Nine sets of "two kangaroos busy cooking dinner in a kitchen" 🙂Generated by Make-A-Video.(Montage courtesy Yaniv; This kangaroo example had become our go-to example in the last few days to the deadline :))#MetaAIMakes pic.twitter.com/dd8H8m7hi6October 2, 2022
#MetaAIMakes https://t.co/Nd6ZqIv2GY pic.twitter.com/vny4Xq3h8kSeptember 29, 2022
When life gives you lemons 🍋✨So many great requests, keep them coming with #MetaAIMakes! Our team is jumping into the replies here to share creations for some of their favorite prompts. https://t.co/w0RS3i6AH4 pic.twitter.com/V3dH5Ky08eSeptember 30, 2022
How does Meta AI's Make-A-Video work?
Meta AI hasn't actually done anything vastly different from what creators of the current wave of AI still image generators have done. It's used the same diffusion technique, through which the AI model buildings images by denoising virtual static to move toward the desired prompt.
It's only trained the model on captions for still images too, Meta AI says, “a model that has only seen text describing images is surprisingly effective at generating short videos.” However, it also gave the model unsupervised training on unlabeled video content so that it knows what sequential video frames look like – apparently it wasn't necessary to specifically train it to know how to combine them.
When will Make-A-Video be made public?
For now, Make-A-Video is a research experiment, and Meta AI hasn't announced what's plans are. It is, however, inviting people to sign up to join a list to receive new of "any future releases" of its Make-A-Video research.
Things are moving very quickly in the world of AI image generation, it seems logical that Meta won't want to be left out. After all, this is the company that wants to sell us virtual designer clothes for our metaverse avatars (what, you don't have a metaverse avatar?). We wouldn't be surprised then if Meta opens some form of public access sooner rather than later. Open AI launched DALL-E 2 with heavily restricted access in April and just five months later has opened access to all.