Seed 2.0 is one of the most profitable video models in China in the past six months.
With its significantly leading generation effect, it almost redefines the upper limit of AI video creation. In half a year, it was difficult for us to see real challengers.
Until last week.
Seedance 2.5 was released, and models such as MiniMax H3 and Wan 3.0 began to accelerate to catch up with Seedance 2.0.
If you only look at the model ability, Seedance 2.5 is still the strongest one at the moment.
But after I actually experienced it, I had an interesting feeling:
The stronger the model, the more we cannot rely solely on the model itself.
Because Seed 2.5 has begun to move from an "AI video tool that everyone can play" to a "professional video production tool."
This means that it increasingly requires the cooperation of the application layer.
Seed 2.5, comprehensive upgrade of professional capabilities
Compared to Seed 2.0, the upgrade to Seed 2.5 is professional and comprehensive.
Among them, I think there are three most important ones:
First, the video length has been increased from 15 seconds to 30 seconds.
Doubling the length of the video means that the model can carry a more complete narrative and make the actions and plots in a shot more coherent.
Second, the number of reference materials has been increased from the original limit to a maximum of 50.
This means that the model can understand more characters, scenes, objects and styles at the same time, greatly improving the controllability of complex creations.
Third, prices have increased by 50%.
Yes, this is also a very important upgrade.
Because the stronger the model, the more we are willing to use it for more complex content; and the longer the video and the more reference materials, the cost of each generation will increase accordingly.
When I seriously experienced Seedance 2.5 for the first time last weekend, I even had the feeling of "Pig pigs eating ginseng fruits."
I charged 400 yuan and finally generated less than 10 videos in total, of which only one was really satisfied.
The money just disappeared on the spot.
At that moment I began to realize:
Seed 2.5 Pro is no longer a model that "enters a few words, then draws cards and other results."
It is more like a set of capabilities for the professional film and television production process.
To really use it well, in addition to the model itself being strong enough, the application layer also needs to provide cooperation in the material management, modification, editing and creation processes.
Because of this upgrade, it also brought three new problems.
1. Models are more expensive
Seed 2.5 is 50% higher than 2.0.
As the cost of generating a video becomes higher and higher, we have to start calculating:
Is this second of video worth generating?
2. The video is longer, but the modifications are more expensive
A 30-second video is certainly more suitable for narrative than a 15-second video.
But problems also arise:
What if there is only 2 seconds in the 30-second video?
Regenerated for 30 seconds, and the cost directly doubles.
Editing yourself means a lot of extra time costs.
3. There are more reference materials, but it is also more troublesome to organize the materials
Up to 50 reference materials, which sounds very cool.
But if you really want to use it, you will find:
Dozens of materials need to be screened, screened, classified, named, and then uploaded to the model.
This process is repeated every time you create something, which is actually very boring.
So, when the Seed 2.5 API was officially released, I paid more attention to how third-party products would catch this model.
Take LibTV for example.
One of the things it does is interesting:
Instead of simply bringing in Seed 2.5, a creative process was redesigned around model capabilities.
Retake of clips: Solving the modification problem of 30-second video
I used Seed 2.5 to generate a video of a morphing robot cat.
This kind of one-shot video is actually very suitable for a 30-second length.
Actions can occur continuously, the camera does not need to be switched frequently, and the overall look will be more complete than 15 seconds.
But when checking the footage, I found a problem:
A small Bug appeared in the middle of the 30-second video.
If it is regenerated, it will be rebuilt in 30 seconds, which is a very waste of points.
If you edit it yourself, it will take a lot of time to process it.
At this time, LibTV's "clip retakes" are very practical.
It can directly select the clip in the video that needs to be modified, regenerate only this small clip, and then automatically replace it back to the original video.
It originally required to recreate the entire 30-second video, but now it only takes about one-tenth of the points.
And even editing is saved.
The model is responsible for generating it, and the product is responsible for reducing the cost of modification.
This may be the problem that needs to be solved after long videos actually enter the production process.
Wonderful slide + smart reference: 50 materials have finally become easy to use
I also ran a more interesting test.
I want to convert a video of the "Original God" Winter Concert into a "Robot Cat Concert."
In the past, I would have needed:
Screenshot → save → classify → upload → mark → hand over to the model.
When the model can support dozens of reference materials at a time, the process will only become more cumbersome.
But in LibTV, I can directly use the "Wonderful Tab".
Throw in a complete video and it will automatically break down the video into multiple node materials.
For example, this time, it directly helped me remove 12 nodes.
These materials are arranged on a free canvas in groups and can be deleted, added and reorganized at any time.
As a result, the original matter of "sorting out reference materials" became a very intuitive visual operation.
Click directly on which material you need.
If you are too lazy to find it, you can also use "smart reference" and let AI automatically help you match the right reference material.
This actually solves a very hidden problem with Seed 2.5:
Although the model supports 50 reference materials, the real difficulty is not "can I put 50", but "which 50 should I put?"
If the application layer can help users complete this step, the capabilities of the model will be truly released.
In the end, I quickly made a rough version of the robot cat concert.
Next, you can make local adjustments and editing.
Smart editing: You can finally use Chat to cut videos
Seed 2.5 has stronger command-following capabilities, which also means that it has higher requirements for prompt words.
If you directly command the model, you often need to use a lot of professional video description terms.
LibTV's "smart editing" provides another way:
You don't need to study professional instructions, just talk directly to the Agent.
For example, I want to make a commercial for riding glasses.
I first uploaded two product material maps and asked it to generate several sets of short advertising videos based on these materials.
After the first round was generated, I continued to tell it:
I also need usage scenarios such as running, iron three, and extreme skateboarding.
It can continue to generate corresponding advertising material.
In this way, mass production of a set of commercial advertising materials becomes very simple.
In the past, it may have been necessary for a team to complete separately:
Script → lens splitting → material sorting → video generation → editing → modification.
Now many links can be handed over to agents through dialogue.
This ability is particularly valuable for e-commerce and brand marketing.
Because what advertising really needs is often not "a video", but:
A set of videos.
Different groups of people, different scenarios, different selling points, and different platforms all require different versions.
After AI videos truly enter commercial production, the value of batch generation may be much greater than that of single video generation.
Ultra-long video: 5-minute video generated with one click
However, the feature that surprised me the most was probably "ultra-long video generation."
It's not simply letting the model generate video for five minutes at a time.
Instead, he added:
Script writing → Smart reference → Smart generation → Smart editing
The whole process was strung together.
I asked it to try to generate a 2-minute animated short film of "Odyssey".
The whole process took about ten minutes.
I went specifically to see what it was doing at every step.
It turned out that it didn't simply "generate a video."
Instead:
First write the complete script → split the shots → generate each scene → discover the missing shots → automatically supplement them → edit them → finally output them into a piece.
Even after two scenes are discovered to be missing in the middle, it will automatically replenish them and then re-edit them.
In the end, I got a full animated short film lasting nearly 2 minutes.
As a first edition generated by a "card extraction", I think the completion degree is already quite high.
Of course, if it is really to be released, there are still many details that need to be adjusted.
But at least it already has:
A complete story, a basically unified painting style, and a piece of structure that can be continued to be processed.
These things are precisely what AI video tools were difficult to solve in one go in the past.
Another problem: expensive
Three issues were mentioned above, but now there is only the last one:
Seed 2.5 is really expensive.
The official selling price is about 1.5 yuan/second.
At this price, a 30-second video is 45 yuan.
If the entire video is regenerated with every modification, the cost can easily get out of control.
However, LibTV membership packages have limited time discounts. Some plans can reduce the cost per second to far below the official price. The highest-end annual plan can even reach about 0.42 yuan/second.
Calculated at this price:
The cost of generating a 30-second video can be reduced from 45 yuan to about 12.6 yuan.
At this time, the above-mentioned "retake of the clip" becomes even more important.
Because what really determines the cost of video production is not just the unit price of the model.
Instead:
Model price × generation times × modification cost.
No matter how strong a model is, if it takes 30 seconds to regenerate it every time it is modified, the final creation cost will still be very high.
The stronger the model, the more a good product is needed
After testing Seeds 2.5 and LibTV, my biggest feeling was not:
"How strong is Seed 2.5?"
Instead:
For professional video creation, in addition to the model being strong enough, the interaction must also be smooth enough.
This is a very interesting change.
In the past, we always felt that AI would eventually make creation easier and easier:
Enter a sentence → get a video.
But the stronger the model, the more I felt that this was not the case.
Because the stronger the model, the more variables it can handle.
Lens, characters, actions, scenes, materials, rhythm, editing...
Really excellent creation still requires people to make decisions.
Agents can help you perform a lot of work, but as long as the final content is still available to others, the creative results still need to be checked by others.
So, AI video is entering a new stage:
It is not "the model replaces the creator", but "the model + Agent + product + person" works together to complete the creation.
Models are responsible for providing capabilities.
The Agent is responsible for executing the process.
Products are responsible for reducing operating costs.
People are responsible for aesthetics, judgment and final decision-making.
In a sense, the stronger the model, the higher the requirements for the product.
Because the modeling ability itself is just an amplifier.
What really determines whether ordinary users can make good use of it is whether the product amplifies this ability.
This may also be the most noteworthy thing after the release of the Seed 2.5 API.