The generative AI arms race is pretty much underway, and Luma Labs is one of the biggest players thanks to its Dream Machine video generation model.

We've been relatively impressed with the results so far, but Jiaming Song, chief scientist at LumaLabs, predicts it could be the next step.

Speaking to Anjney Midha in an interview shared on X, Song explained that real-time video generation is closer than ever, and Luma Labs' dream machine can move the viewpoint while maintaining consistency between shots.

This shifting of viewpoints, which is not possible with the current "one-shot" state of AI video generation, will allow for more control over how the video is finished, making it a more useful tool in traditional film production.

As Mida explains, AI video generation needs to show that it "does more than actually generate cool frames." As an example, Son points out that traditional models act like "image animators."

As a first example, Song shares a prompt asking Luma Labs to generate a video of a small animated character.

We have seen this technique add additional perspectives and animations before, but this video features cuts and transitions where the camera shifts to a completely different perspective while maintaining knowledge of the subject and its surroundings.

This is one of the key features of OpenAI Sora that excited people when it was first unveiled to the world in February, partly because of its longer generation.

In another image-to-video prompt, a girl stares at a giant eye painted on a wall (which Song says "may look a little disturbing in the first frame"). The given image is the eyes staring at the girl, but the dream machine is able to create a stunned expression on her face while keeping her blue dress and short hair consistent from shot to shot.

Song suggests that this "cause and effect" shows that Luma Labs' video model is adapting to a new level of understanding, being able to consider the human psychology in the situation.