I'm here at the Volcanic Engine press conference, where a dizzying array of products were released.
Doubao large model 1.6, Doubao video generation model Seedance 1.0 Pro, voice podcast model, and end-to-end speech model, etc.
Volcano is still Volcano; truly, the reserves are simply too vast.
Most of them have been covered before, such as the voice podcast model, which is the underlying technology for the KOUZI AI podcast I wrote about a few days ago.
But this time, I think the newest and coolest one is the video generation model Seedance 1.0 Pro.
This thing topped the charts a few days ago. Although the name is different, it's actually the same product.
As soon as the ranking was released, many friends immediately couldn't sit still.
Many friends in finance also came to ask me immediately.
But there's nothing I could say; anything I said would be a leak.
Actually, last weekend, I got early access to this model, which is Jimeng AI Video 3.0 Pro.
I've been having a blast with it for days.
The previously popular Jimeng AI Video 3.0 was actually Seedance 1.0 Lite. You can see Han Qing's review of Jimeng AI Video 3.0 here: First-Hand Review of Jimeng AI Video 3.0: All-Round Quality Improvement, a Hexagonal Warrior with Extreme Cost-Effectiveness.
And this time, I tested a bunch of cases with Jimeng AI Video 3.0 Pro, which is Seedance 1.0 Pro. My conclusion first:
Just like Han Qing's evaluation, it's an even more well-rounded and pure hexagonal warrior.
I'll also release my review, hoping to give you some objective understanding of Seedance 1.0 Pro.
Without further ado, let's begin.
This review is divided into the following dimensions:
1. Multi-shot Combination
2. Motion Quality
3. Emotional Performance
4. Camera Movement
5. Physical Dynamic Effects
6. Stylization
Let's go through them one by one.
I. Multi-shot Combination
This can be considered a consistent feature of ByteDance's video models; you can directly switch scenes within the video.
For example, I have this image.
I wrote a Prompt for it:
A lion in a velvet suit sitting in a convertible classic car, the camera slowly approaching from a low front angle. He sits steadily in the driver's seat, head slightly turned towards the camera, mane blowing in the wind, strong sunlight, sunglasses reflecting cloud shadows and wasteland scenery. He remains motionless, as if waiting for a signal.
Camera switch.
The camera switches to an overhead shot inside the car. The lion slowly removes his sunglasses with a raised paw, eyes looking directly at the camera, fingers tapping the steering wheel, distant engine sounds in the background. He lightly sips his lips, slowly turns his head to look at the far end of the highway, and softly speaks a line: "They're finally here."
Camera switch.
The camera pulls back to a low rear tracking shot. The car starts, exhaust fumes spewing out, he slowly drives away from the camera, his back receding into the distance, the clouds ahead lowering, the sky suddenly changing. The camera finally freezes on a road sign: WELCOME BACK, KING.
Here you can actually see that I used "camera switch" twice as a trigger word. When you write it in, you can directly switch scenes in the video.
Let's see the generation effect of Seedance 1.0 Pro.
This semantic understanding ability is a bit outrageous. Almost everything I wrote in the Prompt was realized within these 10 seconds, and the most outrageous part is this:
At the end, I wrote that the camera finally freezes on a road sign: WELCOME BACK, KING.
I know the text is slightly wrong, a bit garbled, but it doesn't hinder its understanding of my words. This road sign was genuinely generated for me, and the text was truly attempted, even if not as perfectly accurate as Jimeng Image 3.0. But I believe, given time to ByteDance, this won't be an issue.
There's also a cat I really like.
Plus the Prompt:
An orange cat sits on a golden carpet, slowly opening its eyes, eyelashes trembling slightly, the camera slowly pushing forward. Camera switch. Close-up shot, the cat lifts its paw and presses a brick at the edge of the carpet, and the floor mechanism clicks. Camera switch. Full view shot, surrounding candles extinguish simultaneously, the stone wall behind slowly opens, and a bright light shines in.
Absolutely perfect. Text-to-video is also possible. I used a Prompt from Master Zang:
A series of rapidly changing dynamic shots: athletes running under the scorching sun, sweating profusely, drops of sweat falling from their foreheads; surfers riding the waves; a group of young people excitedly jumping at an outdoor music festival. Close-up shots show chilled drinks being opened, bubbles rising. Finally, several people raise their glasses in a toast, their faces radiating satisfaction and vitality.
II. Motion Quality
This time, Seedance 1.0 Pro's motion quality is also top-tier.
First up is Britain's famous tough guy, Bond.
The prompt is very simple: A man aims at the target, raises his gun, and fires.
Racking the slide, raising the gun, aiming, shooting – a set of actions that flow very smoothly.
The recoil at the moment of firing and the muzzle flash reflecting on his face are also very realistic. This part actually falls under the scope of physical law evaluation, but it's a strength, so I'll highlight it first.
Then there's this very abstract skeleton doing tap dance.
Although it's just a skeleton, the range of motion is quite large and powerful, even if this dance is comparable to my own.
Looking closely, this guy is quite impressive; nothing breaks down anywhere.
There are also two guys eating jianbing together, and if you didn't know, you might even think it's from a certain Avengers movie.
And the most difficult, sports.
Prompt: A man running and dribbling, shooting, camera following the man.
Within ten seconds, whether it was dribbling or running, there were no errors, very stable.
The only thing to complain about is that the shot didn't go in. But at least it complies with physical laws, unlike some AIs that use who-knows-how-many dark arts to get the ball in, making Newton's coffin lid unable to stay shut.
Then playing soccer.
Prompt: A player skillfully dribbles past opponents, actions fluid, camera following the person.
Dribbling past opponents was not very obvious; the blocking person just blurred in the foreground. But other than that, the athlete's movements were very stable.
III. Emotion
The most important part here is to let everyone experience it immersively, so I'll show more cases and say less.
One of my favorite shots: running and then crying, I can relate.
A girl looks at the camera and smiles.
A pensive child looking out the car window.
Fear, pupils dilating.
A girl sheds tears.
A boxer gets knocked down but defiantly stands up again.
A curious beagle.
What impressed me the most was this case: I told the model that the astronaut was running out of oxygen, Earth was right in front of him, but he couldn't go back.
These are the two ways Seedance 1.0 Pro showed me:
The first one is very restrained, with no major expressions. A slight smile at the corner of his mouth, looking like he's recalling a memorable experience from his life, or perhaps hazy due to lack of oxygen, about to die.
The second one is an immersive experience of what it's like to be out of breath. Heavy breathing, full of will to survive. The camera cuts, and outside the window, Earth is right there, just a breath away. How could he not be anxious? I feel anxious for him.
Seriously, I wonder which AI performance can win an Oscar.
IV. Camera Movement
Camera movement was already touched upon in the first two sections, but here, we're making it purer (and more showy).
A 360-degree rotation.
Another rotation.
An aerial shot.
And then a car chase.
The stability is so good, it feels like Seedance 1.0 Pro could be used as a drone.
V. Physical Dynamic Effects
This part mainly tests whether Seedance 1.0 Pro can keep Newton's coffin lid shut.
This video involves elements such as horses running, steampunk gears rotating, water splashing, and hair fluttering.
Each element, if singled out, could easily cause problems.
But in this video, except for the slightly stiff mane on the horse's back, I can't find any other faults.
The physical laws on Earth are too simple, let's add some difficulty: space physics.
Not bad, Newton has no complaints.
One underwater.
The floating of hair and clothes, underwater bubbles, and water ripples all conform well to real-world laws.
Applying lipstick, the skin tension is very realistic.
Riding a motorcycle is also very smooth.
Including time-lapse photography effects.
The scene of making pottery together, often seen in romance movies, can now be created.
And what's funny is that the most important part of making pottery isn't the pottery itself, but the physical contact, and these two people's hands never let go from beginning to end.
VI. Stylization
Jimeng's stylistic consistency has always been, in my opinion, the best, bar none.
Here's the consistency effect under a specific style:
The man puts down his gun, pulls out a piece of bread, and starts eating it.
Compared to before, Bond was a true tough guy, while this young man carries a different emotion, like a child lost on his first battlefield.
It's truly, very nuanced.
And the pixelization I did in a previous short video, only Seedance 1.0 Pro can roll out well.
Anime style: hands constantly struggling in water. Camera switch, close-up, the protagonist's fearful eyes.
Two illustration styles with very distinct characteristics.
Concluding Remarks
Above, I believe after reading, everyone will have a clearer understanding of Seedance 1.0 Pro.
It can be said that Seedance 1.0 Pro, this new top-ranked model, truly lives up to its reputation. It has no weaknesses in character actions, expressions and emotions, physical laws, camera movement capabilities, stylistic consistency, and semantic understanding, all of which are at a leading level in the first tier.
Moreover, in sports movements, expression and emotion processing, and stylistic consistency, it often delivers surprises.
It feels like Seedance 1.0 Pro will be dominating the rankings for some time.
Of course, other competitors won't be idle; they are all eyeing it enviously.
AI video is indeed becoming increasingly competitive.
Ultimately, all this competition benefits us, the users.
Currently, Volcanic Engine has also opened Seedance 1.0 Pro to enterprise users. The price for approximately 5 seconds of 1080P video is 3.67 yuan.
It will also be fully launched today on the Doubao App. Open the Doubao App dialogue box, select "Animate photos," enter text commands or upload images, and you can experience it.
So, competition is good!
In fact, I'm quite moved.
As someone who has been playing with AI video since the Runway era, I've seen many excellent AI video products in the past two years. Some were famous for a while but gradually fell behind.
Some silently kept catching up. Some were astounding from their debut and are still climbing to new heights.
I hope to often see the names of domestic models on the rankings.
My sincere wish:
Prosperity for the nation.
The above. Since you've read this far, if you think it's good, please give it a like, "looking good," and share it. If you want to receive pushes first, you can also give me a star⭐~ Thank you for reading my article, and we'll see you next time.
>/ Author: Kazak, Shuishan
>/ For submissions or tips, please contact: wzglyay@virxact.com