You probably have a I’ve noticed some great funny AI-generated videos in recent weeks. Harry Potter Reimagined as a Balenciaga commercial and terrifying footage of Will Smith eating spaghetti, it recently went viral. They highlight how quickly AI’s ability to create video is advancing, as well as how difficult some of the technology’s uses are.
These videos remind me of the moment when AI image-making tools exploded last year, when programs like Craiyon (formerly DALL-E Mini) allowed anyone to conjure up recognizable, if crude and often surreal images, like baby monitor footage Gas station robberies, Darth Vadar courtroom graffiti, and Elon Musk eating crayons.
Craiyon was an open source emulator of the carefully curated DALL-E 2 image generator from OpenAI, the company behind ChatGPT. The tool was the first to demonstrate the ability of artificial intelligence to direct text and transform it into what looks like real pictures and human-drawn illustrations. Since then, DALL-E has been open to all, and programs like Midjourney and Dream Studio have developed and honed similar tools, making it relatively easy to craft complex, realistic images with just a few keyboard clicks.
As engineers tweak the algorithmic knobs and levers behind these image generators, add more training data, and drive more GPU chips to run everything, these image-making tools have become incredibly good at faking reality. To take a few examples from a subsite dedicated to bizarre AI images, check out Alex Jones at a gay pride parade or the Ark of the Covenant at a yard sale.
The widespread access to, and sophistication of, this technology is forcing us to rethink how we view images online, as highlighted after AI-generated images purporting to show the arrest of Donald Trump went viral last month. The incident led to Midjourney announcing that it would no longer offer a free trial of its service – a solution that might deter some cheap bad actors but leaves the broader problem untouched.
As WIRED’s Amanda Hoover wrote this week, algorithms still struggle to create compelling video from a prompt. Generating many individual frames is computationally expensive, and as today’s jittery and choppy videos show, it’s hard for algorithms to maintain enough consistency between them to produce a video that makes sense.
However, AI tools are becoming more adept at editing videos. Mimi Balenciaga, along with reference versions friends And Too bad, by combining a few different AI tools, first to create static images and then add simple animation effects. But the end result is still impressive.
Runway ML, a startup developing AI tools for professional photo and video creation and editing, this week launched a new, more powerful technology for applying stylistic changes to videos. I used it to create these dreamlike snapshots of my cat, Leona, walking through “cloudcape” from an existing video in just a few minutes.
Video: Will Knight/Runway
Various machine learning techniques open up new possibilities. For example, a company called Luma AI uses a technology known as neural radiation fields to turn 2D images into detailed 3D scenes. Insert a few screenshots into the company’s app, and you’ll have a fully interactive 3D scene to play with.
These clips indicate that we are at an inflection point for the AI video industry. As with AI image creation, the surge in memes could be followed by dramatic improvements in the quality and controllability of AI videos that bring the technology to all kinds of places. Artificial intelligence may become an inspiration for some authors. Runway tools have been used by visual effects artists working on the Academy Award Everything is everywhere at once. Darren Aronofsky, Director WhaleAnd black SwanAnd bye he is too Fan of Runway.
But you only need to look at how far advanced images from Midjourney and Dream Studio are now to get a sense of where AI video is headed – and how difficult it is to distinguish between real and fake clips. Of course, people can already process videos with current technology, but it’s still relatively expensive and hard to achieve.
Rapid advances in generative AI can be dangerous in an era when social media has been weaponized and deepfakes have become a propaganda game. As Jason Parham wrote for WIRED this week, we also need to think seriously about how generative AI can restore and repurpose ugly stereotypes.
For now, the instinct of trust in videos can be relied upon, but it may not be long before the footage we see becomes less solid and honest than it used to be.