Google’s new video generation AI model Lumiere uses a new diffusion model called Space-Time-U-Net, or STUNet, that figures out where things are in a video (space) and how they simultaneously move and change (time). Ars Technica reports this method lets Lumiere create the video in one process instead of putting smaller still frames together.
Lumiere starts with creating a base frame from the prompt. Then, it uses the STUNet framework to begin approximating where objects within that frame will move to create more frames that flow into each other, creating the appearance of seamless motion. Lumiere also generates 80 frames compared to 25 frames from Stable Video Diffusion.
Admittedly, I am more of a text reporter than a video person, but the sizzle reel Google published, along with a pre-print scientific paper, shows that AI video generation and editing tools have gone from uncanny valley to near realistic in just a few years. It also establishes Google’s tech in the space already occupied by competitors like Runway, Stable Video Diffusion, or Meta’s Emu. Runway, one of the first mass-market text-to-video platforms, released Runway Gen-2 in March last year and has started to offer more realistic-looking videos. Runway videos also have a hard time portraying movement.
Google was kind enough to put clips and prompts on the Lumiere site, which let me put the same prompts through Runway for comparison. Here are the results:
Yes, some of the clips presented have a touch of artificiality, especially if you look closely at skin texture or if the scene is more atmospheric. But look at that turtle! It moves like a turtle actually would in water! It looks like a real turtle! I sent the Lumiere intro video to a friend who is a professional video editor. While she pointed out that “you can clearly tell it’s not entirely real,” she thought it was impressive that if I hadn’t told her it was AI, she would think it was CGI. (She also said: “It’s going to take my job, isn’t it?”)
Other models stitch videos together from generated key frames where the movement already happened (think of drawings in a flip book), while STUNet lets Lumiere focus on the movement itself based on where the generated content should be at a given time in the video.
Google has not been a big player in the text-to-video category, but it has slowly released more advanced AI models and leaned into a more multimodal focus. Its Gemini large language model will eventually bring image generation to Bard. Lumiere is not yet available for testing, but it shows Google’s capability to develop an AI video platform that is comparable to — and arguably a bit better than — generally available AI video generators like Runway and Pika. And just a reminder, this was where Google was with AI video two years ago.
Beyond text-to-video generation, Lumiere will also allow for image-to-video generation, stylized generation, which lets users make videos in a specific style, cinemagraphs that animate only a portion of a video, and inpainting to mask out an area of the video to change the color or pattern.
Google’s Lumiere paper, though, noted that “there is a risk of misuse for creating fake or harmful content with our technology, and we believe that it is crucial to develop and apply tools for detecting biases and malicious use cases to ensure a safe and fair use.” The paper’s authors didn’t explain how this can be achieved.
