One thing to say, buddy really doesn't know what's going on this month.
These manufacturers are like reading the same yellow calendar, and they are all rushing to do things this month.
On the 12th, we won't say much about the development conference of several companies in a row, and I participated in a bunch of conferences last week, and I saw all kinds of large models on the new ones, and I was tired of shopping for my buddies.
As a result, today, guess what, Byte has officially announced his family's new AI, that is, his bean bag can be generated this time, and the effect is extremely strong.
In other words, after holding it for so long, the bytes that appeared at the end of this month began to be served, and it was a hard dish.
There is no basis for words, so let's show you the effect directly:
For example, in this official demonstration of the "chic and chic scene of Hongchen as a companion", the expressiveness of many people and complex expressions in this video can be said to be quite natural, although I am an old fan of Mr. Martin, I can only say that this game of power is quite heavy.
That is, this thing is only 10 seconds and slow, otherwise it would really be deceived as a movie clip.
There is also this section, the man riding a rocket rushes to the auditorium of the world's highest city and causes a big explosion, this storyboard, the consistency of the painting style of the front and back shots is very coherent; The shot of the man in the middle closing his eyes and dying nervously is also very soulful, and the picture is full. . .
But you have to just watch the official video of this thing, it doesn't seem to feel that it's particularly good, sometimes these AI promotional materials are like instant noodles packaging, it looks like one thing, and it's another thing after opening it.
And on the video generation, the p-map fine-tuning is not unprecedented, and there are also video models that show promotional videos, and they don't let everyone use them until now, and they are all played with futures like scallops.
So although after watching these demonstration videos, we think that the bean bag may really have two hits, but in the end, whether the goods are right or not, we still have to give it a try to know.
No, as soon as the press conference ended, Shichao immediately came to the internal test of this PixelDance (pixel jumping) model, and first arranged the traditional reserved items of our AI evaluation, and invited our old employees to show everyone a hot pot.
Feed the model a photo of the hot pot, and then enter the prompt "The dog stands up, picks up the doll next to him and leaves the screen".
According to our previous experience of evaluating some video models with hot pot, within 2 seconds, Brother Guo's body and face will begin to deform, and there have even been cases of turning hot pot into Labrador before.
But guess what?
This time, there is hardly any shaking, frame drops, or flickering distortion.
The action of getting up in the hot pot is done in one go, look carefully at the moment when it drags the rabbit doll over, the shaking of the doll's ears, the dent on the mat when the foot steps down, these details are everyone's products, and the hot pot has to be shouted: Stable, brother!
If it weren't for the flaws in the hair texture of the hot pot in the next few seconds, this video would have scored 9 points.
Next, we tried the light and shadow effects that made many video generation models fail again and again.
Prompt words "Motorcycles are speeding on the road, and the streetscape quickly recedes"
The streetscape changes smoothly, and the light changes are not inconsistent, especially the lights of the building and the reflections on the floor.
If you have to find fault, it's the car coming from the opposite direction, and the picture is not too well controlled.
In the whole classic eating broadcast, the prompt word this time is "The man facing the camera opens his mouth to eat the food on the chopsticks". The description is relatively precise, and it basically just lets the model point where to hit it.
And the PixelDance model did not disappoint us, the chopsticks were very skillful, the food was actually eaten in the mouth, and the face was not deformed by the action of chewing. Even though there was no cue to the group of people in the back, the model still got everyone moving.
That's it.,Isn't this chewing action a bit deliberate?。。。
In fact, at this point, we have already reached a bit of a bottom on the level of the PixelDance model. But in order to make the test more comprehensive, let's try a few more times.
Let's take a look at this sunset taken by Shichao, the prompt word is "In the distant sky, a black dragon is flying, getting closer and closer to the camera", which is a little complicated.
The sky behind, a row of houses, and the black dragon flying towards the camera are basically complete, and the camera will slowly shoot up and follow.
However, the lower left corner of the original photo should be the fence next to the bridge, I don't know if it's because it's too dark, so the model is not recognized, and it is slightly deformed.
There is also this video of a colleague in the editorial department ordering wheat juice on the hour after work, and what shocked Shichao the most was the effect of 8+1 in the cup.
Because of the shaking caused by clinking glasses, and the hand movements that make the liquid tilt to the right.,Good guy.,Dude seems to have really found a model to understand the laws of physics.。。
Moreover, the flexion and extension of the arm and the movement of the joints also seem to be in line with the physiological structure of the human body.
However, there are not no flaws, such as the wine in the hand is drinking, and suddenly it is fed to the mouth of the colleague next door, and the coaster on the desktop turns into a piece of paper without knowing what happened.
I believe that seeing this, everyone has a basic judgment on the PixelDance model of Doubao.
Although it is not yet explosive, it is certainly the first echelon, and the PixelDance model is indeed unique in terms of picture stability and consistency.
However, for Shichao, it is not very unexpected.
In fact, not only us, but many people have high expectations for the video ability of bean bags.
Although it is a little slow, but backed by Douyin, which was born in the video industry, coupled with the investment in the basic model, Doubao wants to rely on these advantages to catch up, which is actually in line with everyone's cognition.
In other words, it is expected that the bean bag is made in itself, not to mention that the things that people take out are obviously the first gear.
However, Byte took out the bean bag video model at this time, which was actually quite timely, just stepping on the node of the development of the video model.
This year alone, in June, Instant Dream AI was unveiled in the AIGC short drama "Sanxingdui: Future Apocalypse", and on July 13, Kuaishou also launched the AIGC short drama "Mountains and Seas";
In the film and television industry, the volume of AI production is also being shouted louder and louder, before there was AI to restore the 27-year-old Jackie Chan's "Legend", and then there was the summer manga movie "Under the Alien", which also means that the video generation model and the film and television industry are in a sense a two-way rush.
Soochow Securities estimates that the potential industry space for domestic AI video may reach more than 580 billion yuan, and in the all-AI mode, the production cost of film and television dramas can be reduced by more than 95% compared with the traditional model.
However, this trend will also bring new problems, such as model training needs more realistic materials, then these video materials may have copyright and privacy issues.
On the other hand, film-level video generation models also require AI manufacturers to have film-related experience, such as this bean bag, which integrates the experience of professional editing and color grading software such as Jianying to make light, shadow, and color effects that are closer to film and television.
Other video generation models may need to be more closely related to the film and television industry if they are to continue to improve in this area.
In short, the volume handed over by this wave of bean bags deserves a high score, but this is far from the end for the video model, and I look forward to the next performance of the domestic large model.
By the way, why hasn't Sora squeaked yet? If you don't give it to it, the daylily will be cold.