How well does Hailuo 02 generate video? Real-world testing shows it surpasses Keling 2.0 and Veo 3.

This Monday, MiniMax quietly announced the start of MiniMax Week late at night, and it turns out that every day is a bombshell!

On the first day, theyopen sourceMiniMax-M1 inference model.

supportWorld's longest context windowThis includes 1 million tokens as input and 80,000 tokens as output.open sourceIn the Open-weight Models benchmark, MiniMax-M1 outperforms DeepSeek in multiple performance metrics!

the next day,Hailuo 02 video model launched.

Tusheng Video has topped the AA Video Arena rankings, outperforming Google Veo 3 and Kuaishou Keling 2.0.

Let's look at this incredibly difficult acrobatic video released by the official team, featuring "knife throwing, jumping through fire hoops, aerial rings, and high-wire walking":

This light and shadow, reality andAIWho can tell the difference anymore?

existAIIn the video industry, efficiency, cost, and quality are like three major obstacles facing developers and creators.

In the pastAIMaking videos isn't something you can solve by spending money; I feel the pinch in my wallet every time I finish a short film...

This MiniMax update is definitely a crucial step towards solving these problems.

1. Cost and efficiency

While significantly improving performance, the Hailuo 02 has almost no increase in training costs, making it the lowest-priced option in the top tier.

Furthermore, the fact that magic is not required already represents a significant increase in efficiency and a reduction in costs.

2. Performance

MiniMax's newly designed DiT architecture improves training and inference efficiency by 2.5 times each, eliminating the need to wait for several hours.

The number of model parameters has increased threefold, the amount of data has increased fourfold, and it has completely broken through the performance ceiling of traditional video models.

3. Instruction understanding

Whether it's camera transitions, lighting changes, or character movements, Hailuo 02 can accurately reproduce them all.

Even with complex commands, its response rate is excellent.

In summary, the Hailuo 02 is fast, cheap, and easy to use...

However impressive the data may seem, we still need to examine the actual performance in practice. Let's put it to the test.

We open the video function in MiniMax and select the Hailuo 02 model.

up to dateThe upgraded Hailuo 02 model has increased its minimum resolution from 720p to 768p. A 6-second video costs 25 points, and the price is the same as Hailuo 01.

The key feature is that it now supports generating 6-second 1080p HD videos, costing only 80 points per use, making it very cost-effective.

Let's take a look at the 768p quality first:

Case 1: Real-world scene reconstruction - characters

Prompt wordsThe girl leaps, flawlessly executing a full rotation in her dive, entering the water with a smooth, silky touch, leaving only a tiny splash on the surface. [Follow, move downwards]

The scene of a girl diving from the platform~Prompt wordsCompareSimpleHowever, the results were quite good, and the various movements were not significantly distorted.

Although there are some imperfections in the athletes' bodies when viewed frame by frame, the overall footage is still quite smooth.

Case 2: Virtual Scene

Prompt wordsThe kitten runs away in terror, holding a camera, while a brown bear chases after it, roaring. [Follow]

The expression and demeanor are so well portrayed; I can actually read what a kitten is thinking from its face! Someone, please save this kitten!

Case 3: Virtual Subject + Special Effects

Prompt wordsThe dragon roared and howled, stirring up a huge mist of water. Suddenly, the dragon charged forward and swallowed the girl whole. The stone shattered and sank under its attack. [Move left, rise]

Environment: Strong winds are blowing all lightweight objects.

Dynamic: Hair and clothing fluttering violently

Background: Changing clouds and lightning in the distant sky.

Light: changes rapidly

Elements: Visible raindrops and water mist

Here, I used the preset vocabulary in the tool to enhance some elements of the environment, background, and screen, and the atmosphere really did look more authentic.

The official documentation also provides many references, such as overall style, camera path, lighting effects, and color tone.

We can also customize preset effects to create sequential shots, which is very convenient when we need consistent characters/scenes/styles.

Try the upgraded 1080p HD version:

Case 4: Expression of Character's Emotions

Prompt wordsThe girl forced a calm smile, glancing around as if avoiding the camera, then looking back at it. Unable to hold back any longer, she lowered her head and wept, her subtle facial expressions conveying an award-winning performance. [Fixed camera angle]

With her downturned lips and slightly furrowed brows, Hailuo 02 has a remarkable grasp of micro-expressions. Even in just 6 seconds, the shot is quite story-driven.

It even perfectly captures the emotional shifts I planned; it's not just about understanding.Prompt wordsIt didn't just read the images in my mind!

This emotional control is already better than many actors, right? I want to see this master use it to make a short drama!

Case 5: Real-world scene reconstruction - animals

Prompt wordsThe little squirrel straightened its upper body, looked around, and suddenly dropped the nut, nimbly running upwards off the screen. [Fixed camera]

The squirrel's fur looks very realistic, and its movements are indistinguishable from a real squirrel.Prompt wordsIt performed a series of complex actions quite well.

The path of the pine cones as they fall also demonstrates excellent adherence to the laws of physics.

Case 6: Effects of a racing drone

Prompt wordsFrom the perspective of a racing drone, it dives low, flies at high speed through a cave, then makes a sharp turn and takes off.

The initial footage shows a thin mist behind the cave, and after passing through the cave, you immediately experience the feeling of being surrounded by clouds and mist.

Hailuo 02 has an exceptional understanding of the environment!

It also designed its own acceleration and deceleration processes, which are quite natural, and the overall visual changes are very smooth.

Case 7: Special Effects

Prompt words[Fixed shot] The fire qilin descends from the sky, crashing explosively onto the shattered street, its four corners hitting the ground. Flames spread around its body, and debris and flames fly everywhere. Skyscrapers collapse, cars...

Finally, I tested the effect with Wensheng Video. Hailuo 02's semantic understanding is truly amazing; it produced this effect in just 4 minutes. The special effects artist was moved to tears…

A few months ago, when I used Hailuo 01 to make martial arts video tutorials, my assessment of it was: "abstract".

The current Hailuo 02 is beyond my reach; sometimes the generated results are even better than I had imagined.

As long as you follow the formula of "clear subject + detailed scene + movement/change (+ camera movement + aesthetic atmosphere)", you can achieve the desired effect.

The value of Hailuo 02 is not just that it's "fast and cheap".

It's about enabling every ordinary person to tell their own story well.

I believe that the future will definitely be an era of "imagination-based thinking".