The AI phenomenon continues its momentum. Building on the success of ChatGPT and DALL-E, OpenAI revealed last February the upcoming arrival of Sora, a video-generating AI, with stunning visuals due to their photorealism, but also their consideration of physics. As was the case with its chat-bot, many companies have followed suit, Luma AI being no exception to this rule, which recently released Luma Dream.
The American firm is throwing itself into the deep end with Luma Dream, all the more cheerfully since the latter can do well in the face of the delay in the release of Sora for security reasons, while it is already available and completely free .
Simple operation…
The concept is simple: like with Midjourney or DALL-E, type a text prompt describing what you want to represent or upload an image. Luma then does so.
Don’t neglect the type of shot, the atmosphere, the protagonists and what they do. You can be more vague and allow the AI the luxury of completing your description. It is also possible to have a photo as a basis and then bring it to life with a new scenario. This functionality has largely enabled the Dream Machine to explode on the networks, with Internet users offering thousands of revisited versions of the most famous memes. On paper, in around 2 minutes, the video is ready.
In reality, it may take longer due to high traffic on the site. To remedy these problems that the sudden craze has generated, Luma AI announced on its networks to multiply the reception capacity by 10. However, even after this scaling, some waiting times are far from the promised 2 minutes. Up to 30 generations of video are free per month. Beyond this limit, you will need to subscribe to a package of $29.99 to $499.99. The more precise you are, the more the rendering will be faithful to what you imagined.
…but which requires a certain rigor
In terms of performance, the company promises on its site that its Dream Machine “understands how people, animals and objects interact with the physical world”. If the promotional or artist-made previews seem enticing, the reality can seem a little brutal: in the majority of cases, the rules of physics are not, or even poorly, respected.
For example, by asking it to represent a boat trapped in a storm, the AI represents the boat, the storm as well as the atmosphere, except that the boat is moving… sideways.
Despite some shortcomings in photorealism, this generator has the capacity to bring to life rather convincing creations, provided that you take care of the description submitted. From a cinematographic point of view, it can help to concretely visualize a shot idea that comes to mind. Like here where Dream Machine was instructed to create a distressing scene around a man in a supermarket.
Even if Dream Machine has some flaws, its free nature and especially its immediate availability make it the spearhead of the democratization of AI, when many companies (OpenAI included) charge to access their full potential. For the moment, Sora’s absence allows him to position himself as a leader in this area, even though we are witnessing its beginnings. Small reminder, Sora, which stunned the Internet with the quality of its one-minute videos, is due to be released in 2024.