After trying the new Runway model, I think OpenAI’s Sora is really dangerous

After trying out Runway's new model, the onlookers cheered and laughed, and the various shots made it hard to tell whether they were virtual or real. Let's take a look at Runway's image, text and video generation capabilities.

Before starting today's article, the Zhiwei editorial department would like to solemnly shout out to Sora: Brother, if you don't open the test, this fan will be ruined. Because today, the highly anticipated Runway finally fully opened the Gen3 version. The length of the generated video of this new generation version has jumped from the original 4s to 10s.

If you casually flip through the comments, you will find that they are all cheering and excited. Some onlookers even said harshly: This is the best video AI now.

Judging from the examples given in this official tweet from Runway, the effects of the Gen3 version are indeed eye-catching. Various camera arrangements, high-definition skin textures, and natural character movements, who can still distinguish between virtual and reality?

Seeing this, one can't help but worry about Sora, which has no signs of opening yet. Because, in the field of live video, everyone is scrambling to release powerful products. From Pika, to Luma, to Kuaishou Keling AI, which has been killing it overseas recently.

Previously, foreigners would even type “Please” in Chinese in order to seek the testing qualification for Kuaishou Keling AI. The full opening of Runway Gen3 this time has drawn everyone’s attention away from Keling.

Since Sora cannot be tested yet, this time the Zhiwei editorial department compared Kuaishou Keling AI, a cultural video model widely recognized both at home and abroad, with Runway Gen3, so that everyone can see the effect of Runway Gen3.

Before the test, let's first explain that since Keling AI is produced by Kuaishou, it is particularly good at pictures related to eating. On the other hand, Runway3 focuses on promoting its sense of reality, imagination, and good lens aesthetics. You can watch it with the characteristics of both.

First, we started with Runway Gen3, a new player’s specialty, and gave it a key sentence that could bring out the cinematic feel: A lonely man drinks a drink in a bar.

( Hints: Static, locked-off shot, Harsh, neon lights, A solitary figure sitting at a bar, Sipping a drink, staring into space, A crowded, dimly lit bar, Lonely, melancholic ) Before the results are out, Runway has a small advantage in terms of generation time, because it takes several minutes to generate content that takes 5 seconds.

But Runway took less than a minute to make a video that was twice as long. And don’t underestimate the short time it takes, because quick work can produce fine results. After the camera zooms in, the man turns sideways and opens his eyes, and a series of camera movements are a surefire million-dollar production. If it weren’t for the unnatural close-up of the open mouth, it would definitely fool many people if they said it was a clip from a Hollywood movie.

Runway Gen3↓

However, as a character that has conquered foreigners, the AI ​​generation effect of Ke Ling is not inferior at all. Moreover, the expression of the protagonist when he looks down at the cup perfectly fits the setting of "Loneliness".

Keling AI ↓

Since there is no winner, let's have another round. This time the key sentence is: the sense of drama of firefighters putting out fire.

(Cue words: Silhouette shot, Fiery flames of a burning building, A firefighter battling the blaze, Spraying water, rescuing people, A burning building in the night, Dramatic, heroic )

In this game, the advantage of Runway Gen3 is particularly obvious. The lens processing of the character silhouette and the orange color make the whole picture solemn and passionate.

Runway Gen3 ↓

Although the generated content of Keling is still fine and the characters' movements are very realistic, the picture processing is a bit random, and it looks like a warm-hearted citizen who accidentally passed by the fire scene and recorded it with his mobile phone. It is very realistic, but the aesthetics of the lens are indeed not as good as Runway.

Keling AI ↓

Since Keling was slightly inferior in the first round, let's switch to Keling's home turf: eating broadcast. However, when testing here, there was a small incident. The Zhiwei editorial department found that Runway has become stronger, but the censorship has also become stricter.

I wanted to try a classic, invite Mr. Will Smith to come out and let him simmer noodles for us again. But I was unfortunately informed that [This content has been marked]. Not only celebrities, but even Pixar animations have been banned.

There is no other way, we can only change the topic to "A man eating noodles".

(Prompt words: Warm, inviting lighting, A person with a focused expression, Using chopsticks or a fork to skillfully slurp up the noodles, A casual dining environment, Enjoyable, satisfying)

The Keling AI, which was at home, performed stably and was invincible. The effect was even more realistic than my first test. I felt the hot steam from the noodles floating on my face.

Keling AI ↓

When I got to Rumway, I made the same mistake as the AI, with the noodles moving instantly and the mouth being misplaced. But the effect was actually much better, at least it knew how to use chopsticks, but it was a pity that I met Ke Ling, the king of eating broadcasts. Runway Gen3 ↓

After trying out the eating show, the Zhiwei editorial department also played a lot of other body movement questions. They even asked them to draw things that don’t exist in reality, such as: let a sheep ride a bicycle.

(Prompt words: Playful POV shot, following the sheep as it rides, Warm, sunny day, A fluffy white sheep perched atop a vintage bicycle, Pedaling enthusiastically, weaving through the countryside, A winding country road lined with wildflowers, Whimsical, joyful)

In fact, the effect on both sides is very good, and both sheep are pedaling very hard.

Runway Gen3 ↓

Keling AI ↓

However, the more the Zhiwei editorial department tested, the more they discovered one thing. That is: no matter how powerful the video AI is, it cannot overcome the real physics engine. For example, I asked them to play basketball for me.

( Hint: Dynamic tracking shot, following the man's movement, High-contrast lighting with a spotlight on the basketball, Subject A young man in stylish black overalls, Dribbling the basketball with flair and precision, An outdoor basketball court with vibrant graffiti art, Energetic ) Runway shows the ball being dribbled in the air and dribbled underfoot. It's hard to tell whether it's basketball or football.

(Playing basketball may be the reason for semantic understanding, because the word "Dribbling" can be used for both football and basketball, and AI may not be able to distinguish and understand it)

Runway Gen3 ↓

When it comes to the Keling AI, the scene is even weirder. The basketball that was thrown out splits into two parts, and the person also grows a third hand.

Keling AI ↓

Basketball at least ensures that the basic movements are correct, but when it comes to collisions and tumbling, the two sides are highly consistent in abstraction, such as this egg that no one can break. (Cue words: The egg fell from the table and broke on the ground, Slow-motion, close-up shot, Natural light from a window, Suspenseful, anticipation building)

Runway Gen3 ↓

Keling AI ↓

Although there are shortcomings, the overall experience shows that Ruwnay3 is basically better than the popular Keling AI. In most cases, it can achieve the official propaganda: realistic, fast, and stable. With the added bonus of long duration, it can give you a high-quality result no matter how wild your imagination is. It really makes the viewer feel that the boundary between reality and virtuality has blurred a little.

(Prompt words: Navigating through a dense jungle and stumbling upon an ancient, glowing interdimensional portal, science fiction, pov. The air is thick with mystery and the sounds of exotic wildlife. As you approach, the portal pulses with an eerie light, drawing you closer to its otherworldly energy. )

Runway Gen3 ↓

However, don’t rush to cheer for Runway, because the game is not over yet. In addition to the effect, there is another unavoidable problem in the field of AI, which is: commercialization. Zhiwei’s editorial department has previously introduced that according to the calculations of domestic manufacturers, a video of about two minutes will cost at least 180 yuan.

The high consumption of the bottomless pit has caused most of the video generation to be stuck at around 5 seconds. Even many of them have very abstract generation effects, which are completely different from the grand scene after the emergence of large language models. Although Runway Gen3 has broken through the 10s barrier this time, the corresponding cost for consumers has increased.

Compared to the previous generation, we paid almost double the money. If calculated according to the monthly package, the basic membership is $15 per month, and the buyer can only get 625 points. However, Runway3 consumes 10 points for every second of video, so a 10-second video is worth $2.4, which is about RMB 17.

You can calculate how much money it would cost to just use the examples shown in this article. No one can stop this because it is an inevitable trend, and profitability is a problem facing all large AI models.

So, although Runway3 is very strong and won the game, we still have to admit that the completely free Keling is more attractive. As for Sora, we have already begun to feel disenchanted. Before he is no longer useful, he is just a beautiful fantasy bubble, 0 points.