Didn't Shichao just write the function of Runway that can be moved by just applying it two days ago? Unexpectedly, before the excitement has passed, another new product will hit the market. The protagonist this time is a man namedPikavideo generation model. This company was only established in April this year. Even though it is still a fledgling company, it has already raised funds of$55 million, even scientists from OpenAI came out and started the boast mode.
Even the listed company of one of the founder’s fathers was still popular due to Pika.daily limitYes, it’s really only the father who is more valuable than the daughter...
Shichao also smelled it and went directly to their official website to find out.
As a result, their promotional video fooled me when it came out.
it exceptJust throw in a picture and a video will be generatedoutside.
You can also generate a video with just one or two prompt words.
For example, let it generate a 3D animated horse talking in space.
This face shape and demeanor are probably to the extent that even Musk himself would call him awesome when he saw it.
Just drop a video in and Pika can generate new content based on the prompts.
Like in the demonstration, if you put a silhouette video of a young man on horseback into it, the corresponding style can be generated based on the prompt words.
And after splicing together various styles of realistic style, 3D cartoon style, and pixel style, you can switch between themExtremely silky, riding a horse the whole time.There is no deformation in the movements at all.
Even this Pika can directly view videosMake local modificationsYes,Hit wherever you point…
Just drag the mouse lightly to frame the elements to be modified in the video, and you can change the girl's outfit in minutes and put on sunglasses for the chimpanzee.
To put it bluntly, the functions and effects demonstrated above can almost defeat most of the video generation AI.
However, the above official demonstration videos, everyone dddd.
What are the differences between actual and demonstrated? As always, Shichao will give you a review first.
We still have to wait in line for the official web version, so we can only try it in the Discord community first."castrated version", it only has two functions: text generation video and picture generation video.
After opening the Pika community in Discord, the main body of the entire interface is the introduction. There is a "CREATIONS" drop-down entrance, there are 10 generated entrances, you can experience it by entering any one.
It’s very simple to use, just hit"/", the suffix will pop up automatically. The two main functions of Pika are /animate (picture to generate video) and /create (text to generate video).
After selecting the mode, just like using ChatGPT, just type or enter pictures to describe your requirements.
Although this Pika does not have a special parameter adjustment column like Runway, you can still adjust parameters if you want to, just add the suffix.
And there are many parameters that can be adjusted by this Pika. One of the newly released functions this time isAdjust the lensoptions.
You can adjust things like zooming in, zooming out, moving up, down, left, and right, rotating in various directions, just like being a director.
Other parameters include the number of frames, the amplitude of the picture change, the video ratio, etc., and they are all required by adding a "-" after the prompt word.
After understanding the parameters, you can start playing directly. Shichao first tried the function of generating videos from pictures and threw it a picture of snowing. It took about a while.one minuteJust about right.
The effect looks pretty good, but the trial version of Pika only lasts for up to 3 seconds at a time, which feels a bit unfinished.
Next, I will throw away the entire classic meme that I tried with Runway last time. Although Pika's picture is a bit blurry, in terms of the continuity of the movements, it is still slightly better. At least it does not change clothes in the street, and the rhythm of walking is also normal...
Comparison of the effects of Pika (top) and Runway (bottom)
However, there are also times when the car rolls over. This is still the case I tried last time. It made the long-haired beauty below have flowing hair, and the Runway effect was pretty good.
But something is not right about Pika. If her hair didn’t float up, I wouldn’t even know there was a human head hidden in it. A good picture would immediately turn into a ghost movie...
Comparison of the effects of Pika (top) and Runway (bottom)
Shichao also added "-camerarotateclockwise" after the prompt word and tried Pika's new "lens control" parameter to make the pictureRotate clockwisestand up.
It looks pretty good at first glance, but unfortunately these butterflies did not flutter twice when the camera was rotated, and they also turned in the wrong direction.
After trying the function of generating videos from pictures, we tried generating videos from text.
Shichao found that for someSmall movements or simple movementsAccording to the requirements, the effect generated by Pika is not bad.
For example, let Pika generate a 1:1 video of a cat sleeping, which is quite satisfactory.
But as long as it involvesCoordinated movement of limbs,orhand detailsAt that time, that wasFull of flaws.
Let a 3D version of Mario run. Even if the frequencies of the hands and feet are out of sync, why do the legs still switch back and forth, and one hand has four fingers and the other has five fingers...
There was also a dance by Michael Jackson. One second I thought it would work, but the next second he suddenly grew two little arms.
What’s interesting is that I also tried out the prompt words in the official demo, and all I could say wasToo abstract.
Musk’s one doesn’t have the same effect as in the demonstration video at all. Let’s just look at the effect. It’s hard to comment...
There are also prompt words for the robot to walk. The demonstration picture is like this:
But when we tried it, it looked like this. Although the effect was okay, it was not as good as the demonstration.
After trying too many times and finding no better cases, Shichao went through the videos generated by everyone in the community and looked at the official database of excellent cases. It turned out that there was no good effect.It’s really not my personal fault.
Everyone uses Pika to generate good results, but most of them are videos in which some elements of the screen flicker or have little movement range. As long as the movement range is larger, all the shortcomings will be exposed.
The following one is still included in the excellent case library, and the calf side looks a bit strange...
In general, judging from the experience on Discord alone, the effect of Pika is average, similar to the effect last time on Runway, with good and bad results.
However, the official version of Pika will have more functions in the future, and maybe the ones in the demo picture will beUpload a video and modify part of the videofunctions can also be used.
Looking back now, in the past few months, almost all the models for AI-generated videos have been updated together. Runway, StabilityAI, Meta, etc. have all released new models for generating videos.
Although the effects of these AIs currently on the market do not seem to be up to the level of standard videos, as long as you have a little understanding of the development of AI-generated videos, you will find that the recent wave of updates has made great progress.
Like the videos generated by AI a few months ago, it was accompanied byThe background is still changing, and now,The background is much more stable.
This is only the result of just a few months. Maybe in a few months, this wave of models will be updated again, and the effect will be completely different.