You will also be interested
[EN VIDÉO] Interview: How was artificial intelligence born? Artificial intelligence aims to mimic how the human brain works, or at least its logic…
From mid-2022, one topic caused a stir: “Text to Image” applications (AI Image GeneratorAI Image Generator). And for a good reason. We type in some text and get back an image that is mostly of high artistic quality. OpenAI’s Dall.e 2 got the ball rolling but was eclipsed by MidJourney and Stable DiffusionDiffusion – and for our part we think Leonardo.ai is the most powerful of them all.
The next holy grail is text to video, which is creating videos from plain text. Apps like Genmo and Kaiber have taken the lead, but so far they leave a lot to be desired. The goal seems much more ambitious than with “Text to Image”.
It remains that for a few weeks there has been an excitement at the level of the main protagonists of the field. Four giants have stormed this fortress, each hoping to do well and be at the center of artificial intelligence (AI) after the shock caused by OpenAI’s ChatGPT. However, it is a stranger who seems to be furthest along in this quest…
Google Pictures
If there’s one company that can win everything by making its star shine, it’s GoogleGoogle. Over the past ten years, it has become a leader in artificial intelligence, whether it’s with its autonomous vehicle or an AI’s victory over one of the world champions of the game of Go. However, it is an understatement to say that Google was surprised by the arrival of ChatGPTChatGPT, which overnight threatens a rule that it would have thought untouchable.
With Imagen, Google is trying to regain control. The search giant offers us a series of five-second clips that reveal a certain level of know-how, but without being mind-blowing. And the wait is long.
Google Imagen demo: “A panda drives a car”. ©Google
Meta, the company that powers Facebook and Instagram, has a number of slightly more adventurous clips on their Create Video page. But again, we are in no way offered to test the beast. It is true that Meta probably has painful memories of his failed essay on the Metaverse, which had been proposed a bit too soon to netizens’ curiosity.
Nvidia “VideoLDM”
On April 23, Nvidia, the leading graphics card vendor, showed at an Institute of Electrical and Electronics Engineers (IEEE) conference that its own “text-to-video” application, VideoLDM, had made honorable progress. From a short sequence on the freeway to a teddy bear playing guitar, we can see that the technology is advancing (Nvidia is attacking high image resolutions), but their solution still needs improvement. And even then we can’t test the thing yet.
“A cat with glasses guards a swimming pool”. This video was created with Nvidia’s VideoLDM tool. © NVIDIA
Adobe Firefly
Adobe has been positioning itself in the AI space for a few months with the Firefly collection and already offers several high-quality applications, including one that cleans up audio files. In order to be able to test the promising home application “Text to Video”, you have To on a waiting list. We’ve applied and we’re still waiting…
Runway GEN-2
In reality, the surprise seems to be coming from a start-up start-up: Runway. This offers a toolbox with around thirty applications of artificial intelligence. And also, good news, some of these tools are now available on iPhoneiPhone under the RunwayML name. However, Runway has already excelled with a Gen 1 application that can transform an existing video into an amazing creation (Gen 1 is also available on iPhone recently).
Now Runway dangles its successor Gen-2. Based on the know-how this company has already demonstrated in relation to AI and in particular with Gen-1, which has continued to develop over the months, there is good hope that Gen-2 will take the leading position of “Text to video” monopolized “.
In fact, if the application delivers what it promises, we should be entitled to clips with stable fluidity, without illogical effects on the movements of image movements. The bets are open.
The Gen 2 demo shows the progress made by the predecessor Gen 1 since its inception. © RunwayML