Skip to content
Author｜Zichuan
Source｜AI Pioneer
The long-absent Luma AI has finally updated!
Today, Luma AI released their brand new product, Ray2.
According to reports, compared to Ray1.6, Ray2 excels in text semantic understanding and natural coherent actions, with a tenfold increase in computing power and faster generation speed.
However, it currently only supports text-to-video functionality, with plans to introduce image-to-video, video-to-video, and video editing features in the future.
Let’s take a look at how powerful Luma AI’s new video model Ray2 really is.
The official demo released covers various styles and genres of videos, with the most prominent features being the coherence of actions and text understanding capabilities.
It’s the first time I’ve seen someone eating sand so naturally.
Not only is there good coherence, but it’s also very realistic; pouring honey finally looks right.
It even had no trouble cutting steak.
After waking up, even gorillas have started extreme sports. I must say, the athletic performance is really impressive.
Even more exciting than the official demo are the creative works from various experts.
When using Ray2 to generate videos paired withMMAudio-generated sound effects, it feels like a Hollywood blockbuster.
To test its capabilities, we need to put it to the test.
As usual, let’s start testing!
Coincidentally, Vidu 2.0 has just been released, so we will use it along with Keling, the dominant player in the video generation field, to evaluate its capabilities.
Next, we will compare Ray2, Keling, and Vidu 2.0 by inputting the same prompts to see which is stronger.
However, Vidu 2.0 currently only supports text-to-image mode, so I will use the first frame of the video generated by Ray 2.0 as a reference image, along with the same prompt to generate.
As we all know, motion performance has always been a major challenge for video generation models, often resulting in deformation or actions that do not conform to the laws of motion.
First, let’s take a look at their motion performance.
Prompt: A capybara riding a bicycle in a winter park movie shot.
Wow, Vidu is just standing still….
Keling is not much better; its feet aren’t even pedaling, and the bicycle moves on its own. Ray2, however, performs very well in accordance with the laws of physics, although the faster it goes, the more its clarity seems to diminish.
However, purely from the perspective of motion performance, Ray2 is clearly more in line with the laws of physics.
Evaluations naturally cannot do without my favorite anthropomorphism.
Prompt: A kitten practicing karate, punches, kicks, flying kicks, karate, wearing karate uniforms. Karate dojo.
Well…., let’s just say you still have to draw cards.
Camera movement is also a major challenge in video generation; often the camera transitions do not follow the prompts, especially in terms of detail description.
Next, let’s see how they perform in terms of camera transitions.
Prompt: The focus shifts from a close-up of the shooting position to the tip of a sniper rifle barrel in the rain. Zoom lens, film, 35mm.
Vidu’s gun is leaking water, haha.
Camera movement is indeed a bit challenging for Keling; on the contrary, Ray2 better adheres to the semantics, even though the prompt did not mention a person, the final result turned out better.
Overall, Luma AI’s new video modelRay2 has improved, especially in its strong understanding of text commands, being able to comprehend interactions between humans, animals, and objects, creating coherent and physically accurate videos.
However, compared to Keling, the image quality still needs improvement!
I originally thought 2024 would be the most competitive year in the field of video generation, but with Luma AI launching Ray2 in early 2025, it seems that the video generation field will welcome a new arms race!
Scan the code to join the group, and we will guide you to explore ChatGPT, GPT-4, Wenxin Yiyan, Tongyi Qianwen, iFlytek Xinghuo and other AI large models, while learning some AI money-making skills.
