Editor: Peach
Gen-3 Alpha has finally started testing! Netizens who got the qualification for the internal test for the first time have released various explosive demos, and it can be seen that Gen-3 has completely jumped in the quality of generation. However, the flaw that models sometimes fail to understand the physical world still exists.
The long-awaited Gen-3 Alpha is finally open to some super creative collaborators.
The netizens who got the internal test qualification are happy.
Compared to the previous generation, the Gen-3 Alpha has been epic in terms of detail, consistency, and athletic performance!
Others say that Gen-3 gives us an exciting future – it won't be long before we can see movies that are entirely generated by AI.
However, it is undeniable that Gen-3 Alpha still has a certain gap in understanding the physical world.
Take a demo where netizens tested the Gen-3 and Sora, and the off-road vehicle drove fast in steep hills, and the tires were dusty.
It is evident that Gen-3 has generated a scenario where dust only appears in front of the vehicle.
Even the previous Sora was also an ant with only four legs, and was noisy by netizens.
Not only that, Dream Machine, another AI video model that has been more popular in the past few days, has made a high-altitude performance of gymnasts into a horror movie.
The limbs of all the performers, when performing skills such as flipping in the air, are either scattered and deformed, or merged into one.
After testing Gen-3, other netizens found that the same problem existed.
The one below performs directly on an imaginary gymnastics pole.
LeCun's taunt is certainly an occasion for such moments, which he excitedly repeats, "Video generation models don't understand basic physics, let alone the human body."
He affirmed that AI video models will get better and better over time.
"However, a learning system that truly understands physics is not going to be generative. All birds and mammals, know physics better than any video generation system. However, none of them were able to generate detailed videos."
Various demos, netizens play crazy
Two days after its opening, the Gen-3 Alpha demo has been released all over the network, so let's take a look.
Let's see what is the strength of Sora's AI video model today? Where is the gap?
The details of the characters are accurately portrayed, and the performance is more tense
Physics in the water, Gen-3 is still amazing.
In the portrayal of the characters, the tension of the performance, I have to say that Gen-3 is really strong.
These hands playing the piano are so perfect.
In Gen-3, you can also draw a smart little girl with rabbit ears.
The author states that he actually made a mistake in his prompt (a girl with rabbit fur), but Gen-3 understands her intentions well and draws rabbit ears.
AI-generated fashion shows, comparable to the world's four major fashion weeks
There are also AI-generated fashion show performances, which completely break the imagination of aesthetics and can lead the fashion frontier.
Elves and monsters are sci-fi in the future, and Gen-3 imagination is too strong
The "holographic" sprites generated by Gen-3 look a little magical.
In the River Thames in London, a hidden monster emerges from the water.
This scene is reminiscent of a scene from a Godzilla movie.
Another netizen made a sci-fi short film called "2030 - Mission Europa" with Gen-3 and accompanied it with music.
And the space elevator, which looks like that.
Multi-scene conversion, control in place
Netizens use Gen-3 to generate a person's experience of flashing through different stages of his life.
From childhood, to elementary school, middle school, to college, and marriage, it is all presented.
In the following scene transition, Gen-3 controls it very well.
From a castle in Ireland, fly over a futuristic cyberpunk city lined with skyscrapers.
There's also this pixel-style AI video generation, which is completely native.
Video games, one-click generation
Huang has said that in the next 5-10 years, we will see games that are entirely AI-generated!
However, the AI video game generated by Gen-3 is also amazing enough.
Netizens used Gen-3 to generate the 16-bit retro video game 1776.
Diverse fonts
In addition, Gen-3 Alpha is simply amazing in terms of the generation effect of the title animation.
Netizens collected 9 of the most classic cases.
After the waves crashed on the beach, the words Ruway were left behind.
The word Ruway appeared on the big screen of the movie theater, giving people a Marvel-style rush.
Fireworks burst into the sky with the word Ruway.
There are also amazing fonts generated by various elements such as bubbles, fluids, and leaves in the ocean.
Failure cases
In addition to the case of Gen-3 not understanding the physical world given at the beginning, netizens also collected some other failure cases.
For example, the knife below cuts the meat, cutting one slice into many slices. Netizens exclaimed, what a powerful knife.
Gen-3、Sora,谁更强?
Who is better than Gen-3 or Sora?
Netizen Champion used 9 of the same prompts to test the two models separately.
Tip 1: A diver discovers a hidden shipwreck from the future with biochemical sea creatures and advanced alien technology on board.
In terms of the generation effect, Gen-3 has a stronger sense of future technology, and Sora only highlights the appearance of the shipwreck.
Tip 2: Super close-up shot of a 24-year-old woman blinking, standing in a magical moment in Marrakech, motion picture film, 70mm shooting, depth of field, bright colors, cinematic effect
With this hint, Gen-3 and Sora have their own merits, and Gen-3's pupils seem to be more recognizable to the magical city of Marrakech.
However, Gen-3's video does not blink and lacks authenticity.
Tip 3: An alien blends naturally with New York City, paranoid thriller style, 35mm film.
As far as the alien image is concerned, I have to say that Sora is better and more acceptable, and the aliens generated by Gen-3 are like a horror movie.
Tip 4: A stylish woman walks down a street in Tokyo filled with warm neon lights and vivid city signs. She wears a black leather jacket, red dress, and black boots, and carries a black handbag. She wears sunglasses and red lipstick. She walks confidently and casually. The streets are damp and reflective, creating a mirror effect of colored lights. There are many pedestrians walking around the street.
And then there's this classic image generated by Sora – a fashionably dressed woman walking down the streets of Tokyo, and Gen-3 doesn't lose a bit. However, when it comes to the specular effect, Gen-3 does not follow the instructions exactly.
Tip 5: Historical footage of the California Gold Rush.
Gen-3 is a more realistic representation of the California Gold Rush, where people are immersed in the joy of gold rushing, and the background color highlights a greater sense of history.
That is, without contrast, there is no harm.
Tip 6: A young professional product reviewer sits in front of a computer with two displays, in a well-lit video studio, surrounded by gadgets and tech devices. He's holding a movie camera and thinking about his next video content. He is aligned with the focal length and the background is slightly blurred to achieve a cinematic effect.
Gen-3 is richer in terms of character expression. Sora just turned the man's head from side to side, and it didn't reflect the action of focusing.
Tip 7: The camera revolves around a large number of vintage televisions, all of which are playing different programs: 1950s science fiction movies, horror movies, news, snowflake screens, 1970s sitcoms, etc., set in a large museum exhibition hall in New York.
The Gen-3 is a good embodiment of rotation, but the TV has a bit of a "mold-wearing" feel. There is also a snowflake screen, which Gen-3 misses.
However, in the case of Sora, there is no spinning.
Regardless, there is no perfect AI video model.
There is also an additional comparative case, you can feel it for yourself.
Resources:
https://x.com/FinanceYF5/status/1807344173944611157
https://x.com/minchoi/status/1807433108456665282
https://x.com/ylecun/status/1807556244581822643
https://x.com/ProperPrompter/status/1806972255706231221