This Monday, MiniMax quietly announced the start of MiniMax Week late at night, and unexpectedly, every day was a bombshell!
On the first day, theyopen-sourced the reasoning modelMiniMax-M1。
supportingthe world's longest context window, including 1 million tokens input and 80,000 tokens output. In the Open-weight Models benchmark tests, MiniMax-M1 surpassed DeepSeek on multiple performance metrics!
On the second day,Hailuo 02the video model went live。
refreshing the AA Video Arena (global video model arena) leaderboard, with image-to-video crushing Google Veo 3 and Kuaishou Kling 2.0.
Now look at this official ultra-difficult acrobatics video featuring "flying knife throwing through a ring of fire, aerial hoop, and high-wire walking":
With this lighting and shadow, who can still tell reality from AI?
01. Hailuo 02's breakthrough advantages
In the AI video field, efficiency, cost, and quality are like the "three great mountains" standing between developers and creators.
In the past, making AI videos was not something that could be solved just by spending money. After finishing a short film, I would even feel sorry for my wallet...
And after this MiniMax update, it has absolutely taken a key step in the direction of solving these problems.
1. Cost and efficiency
While Hailuo 02's performance improves significantly, its training cost has barely increased, and it has the lowest price among the first tier.
Also, the fact that no magic is needed is already a huge improvement in efficiency and reduction in cost.
2. Performance
MiniMax's newly designed DiT architecture improves training and inference efficiency by 2.5 times each, so you no longer have to wait for several hours.
The number of model parameters has surged 3 times, and the amount of data has increased 4 times, completely breaking through the performance ceiling of traditional video models.
3. Instruction understanding
Whether it is camera switching, lighting changes, or character actions, Hailuo 02 can restore them accurately.
Even for various complex instructions, its response rate is excellent.
Overall, Hailuo 02 is fast, cheap, and easy to use...
However, no matter how good the data looks, we still need to see how it performs in actual use, so let's do a round of real tests together.
02. Real test cases
We open MiniMax's video feature and select the Hailuo 02 model.
The newly upgraded Hailuo 02 model's minimum resolution has been raised from the original 720p to 768p. A 6-second video costs 25 credits, the same price as Hailuo 01.
The key point is that it now supports generating 6-second 1080p HD videos, costing 80 credits per generation, with very high cost-effectiveness.
First, let's look at the 768p effect:
case 1: Real-world scene restoration - people
Prompt: A girl jumps, perfectly completes a one-revolution diving twist, enters the water smoothly, and only a small ring of splashes appears on the water surface [follow, move down]
The scene of the girl diving from the platform~ The prompt is relatively simple, but the result is quite good, and there is no major distortion in the various movements.
Although when viewed frame by frame, the athlete's body has some flaws, overall it still looks fairly smooth.
case 2: Virtual scene
Prompt: A kitten runs in terror while holding the camera, with a brown bear chasing behind it, roaring. [follow]
The presentation of expressions and facial expressions is very good. I can actually read what a kitten is thinking from its face. Who can save the kitten ahhh...
case 3: Virtual subject + special effects
Prompt: The giant dragon roars and bellows, stirring up a massive water mist. Suddenly, the giant dragon charges forward and swallows the girl in one bite, and the stone also shatters and sinks under its attack [pan left, rise]
Environment: Strong wind blows all lightweight objects
Motion: Hair and clothing flutter violently
Background: Changing clouds and lightning in the distant sky
Lighting: 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 image, and the atmosphere really feels more convincing.
The official team also provides many references, such as overall style, camera path, lighting effects, and image color tone.
We can also customize preset effects to create a series of connected shots, which is very convenient when you need consistency in characters/scenes/style.
Let's try the upgraded 1080p high-definition effect:
case 4: Character emotional expression
Prompt: The girl forces a calm smile, avoids the camera and looks around into the distance from side to side, then looks at the camera again, can no longer hold it in, lowers her head and cries aggrievedly, subtle micro-expression changes, award-winning actress-level performance, professional shot [fixed shot]
The downturned corners of the mouth and slightly furrowed brows, Hailuo 02's control of micro-expressions is very impressive. In just a 6-second shot, it already feels quite story-driven.
It can even present the emotional turning point I set very well. It not only understood the prompt, but understood the image in my head!
This emotional control has already surpassed quite a few actors, right? I want to see a big creator use it to make short dramas!
case 5: Realistic scene restoration - animals
Prompt: The little squirrel straightens its upper body, turns its head looking left and right, suddenly, it drops the nut and agilely runs upward out of the frame [fixed shot]
The squirrel's fur still looks very realistic, and its movements and form are no different from a real squirrel. It completed the series of complex actions in the prompt quite well.
Judging from the falling path of the pinecone, it also follows the laws of physics very well.
case 6: FPV drone effect
Prompt: FPV drone perspective, diving in low-altitude flight, speeding through a mountain cave, then making a sharp turn and soaring into the air.
In the initial image, you can see thin mist behind the mountain cave. After passing through the cave, you immediately experience the feeling of being surrounded by clouds and mist.
Hailuo 02's understanding of the environment is top-tier!
It even designed its own acceleration and deceleration process, and it feels quite natural. The overall image changes are very coherent.
case 7: Special effects
Prompt: [fixed shot] The fire qilin descends from the sky, crashes onto the broken street with explosive force, lands on all fours, flames spread around its body, and flying debris and flames burst out all around. Skyscrapers collapse, cars
Finally, I tried the text-to-video effect. Hailuo 02's semantic understanding is really impressive. It produced this effect in 4 minutes. Special effects artists are already crying...
03. Some sharing
A few months ago, when I used Hailuo 01 to make a martial arts video tutorial, my evaluation of it was: "abstract".
But now Hailuo 02 is in a state I can hardly reach. Sometimes the generated effect is even better than what I imagined in my head.
As long as you follow the instruction formula of "clear subject + detailed scene + motion/change (+ camera movement + aesthetic atmosphere)", you can get the ideal effect.
The value of Hailuo 02 is not just "fast and cheap".
Instead, let every ordinary person be able to tell their own story well.
I believe the future will definitely be an era of "imagination as the standard."
Original link:Surpassing Kling 2.0 and Veo 3, the new king of cost-effectiveness among AI video models has arrived!