Name: Seedance 2.0 究竟有多神？為何好萊塢都嚇傻了！ (How Seedance 2.0 is SO GOOD (And Why Hollywood is Shook))
Uploaded: 2026-03-04T14:18:45.000Z
Duration: 10 min 39 s

Well, Sea Dance Fever continues to grip the AI video landscape.

哇，Sea Dance Fever 繼續席捲 AI video 的領域。

So today I'm back with some more information on the model that has rocked the pillars of what we thought was possible.

所以今天我又帶來了更多關於這個模型的新資訊，它徹底顛覆了我們對可能性的想像。

To that I'm digging into some interesting insights on how this model works and well, I mean, what that means for the next wave of updates from other AI video models.

接著，我會深入探討這個模型是如何運作的，以及這對其他 AI video models 未來的更新意味著什麼。

Plus we're going to talk about Sea Dance 3.0.

另外，我們還要聊聊 Sea Dance 3.0。

You have information about Sea Dance 3.0?

你有關於 Sea Dance 3.0 的資訊嗎？

Welcome to my third video on Sea Dance 2.0, a video model that has not been globally released yet.

歡迎來到我第三支關於 Sea Dance 2.0 的影片，這是一個還沒全球發表的 video model。

I mean, obviously the hype is pretty high on this one.

我的意思是，這東西的討論度當然很高。

Now as a quick note, I do have access to the model and I have been testing it, but

順帶一提，我確實有權限可以使用這個 model，而且我已經測試過了，但是

you know, because of early access agreements, I'm actually not allowed to post my results yet.

你知道的，因為有 early access agreements，我其實還不能發布我的測試結果。

That said, we do have plenty of other examples to look at either by those in different early access programs or those generating on the Chinese platform.

話雖如此，我們還是有很多其他的例子可以看，不管是來自其他 early access programs 的人，或是那些在 Chinese platform 上生成的人。

But honestly, this video is not a collection critiquing the video outputs of the model, although, I mean, I can say it is really good.

但說真的，這支影片不是在評論這個 model 的 video outputs，雖然，我的意思是，我可以說它真的很棒。

Instead, today I wanted to take a look at why the model is so good.

取而代之的是，今天我想來看看為什麼這個 model 會這麼好。

And well, to answer that, we have to look at seed 2.0.

嗯，要回答這個問題，我們得看看 seed 2.0。

But in terms of its performance, I mean it is considered on par with flagship models like GPT 5.2 and Gemini 3 Pro.

但就它的表現來說，我認為它已經可以媲美 GPT 5.2 和 Gemini 3 Pro 這類的旗艦級 model 了。

But the really important part here is that C 2.0 is doing this at a cost 10 times cheaper than than Google or open AI.

但這裡真正重要的部分是，Sea Dance 2.0 的成本竟然比 Google 或 open AI 低了十倍。

Now you might be saying, well, that's the large language model.

你可能會說，嗯，那是 large language model。

What does that have to do with video generation?

那跟 video generation 有什麼關係呢？

We're going to circle back to it in just a minute.

Now, kicking briefly over to the Google side, we do know that Gemini 3 is like kind of the brains of Nano Banana Pro.

接著，簡單來說一下 Google 的部分，我們知道 Gemini 3 就像是 Nano Banana Pro 的核心。

Or like to put it more accurately, Nano banana Pro is Gemini 3.

或者更精確地說，Nano Banana Pro 就是 Gemini 3。

I mean, technically the name is Gemini Pro Image Preview.

其實，技術上來說它的名字是 Gemini Pro Image Preview。

But well, I mean, let's face it, Nano Banana is a lot more fun to say.

但說真的，我們承認 Nano Banana 唸起來有趣多了。

And that's why you can prompt for things in Nano Banana like when you.

這也是為什麼你可以在 Nano Banana 裡輸入指令，像是當你。

Well, I'm giving it the URL for the Gemini Wikipedia page and then you know, saying, create an infographic of this Wikipedia page and it can do so accurately because it can access and summarize the information it like it reads.

好，我給它 Gemini Wikipedia 頁面的網址，然後說，請為這個 Wikipedia 頁面製作一張資訊圖表，它就能準確地做到，因為它可以存取並總結它讀到的資訊。

Also, I might have just made Nano Banana self aware.

另外，我可能剛剛讓 Nano Banana 產生了自我意識。

And that concept is the key to understanding what's happening with Sea Dance 2.0.

而這個概念，是理解 Sea Dance 2.0 正在發生什麼事的關鍵。

Or I should say at least like this is very highly speculated since they have not actually released the research paper on it.

不過我應該說，這至少是大家高度猜測的，因為他們還沒有真正發表相關的研究論文。

雖然我們可以合理地支持這一點。

I mean, there is no question that C 2.0 is under the hood of Sea Dance 2.0.

我的意思是，毫無疑問，C 2.0 就在 Sea Dance 2.0 的核心裡。

So this is a thinking video model, which you just sit on that for a second.

所以這是一個思考型的 video model，你可以先仔細想想。

Now I'm not going to get overly technical here.

And again, this is extrapolating what we know about Sea Dance 2.0 as well as what we know about Sea Dance 5.0, the image model.

而且再次強調，這是根據我們對 Sea Dance 2.0 以及 Sea Dance 5.0（圖像模型）的瞭解來推斷的。

And that's where we do have some insights into how this model works, as they describe it as a unified multimodal generation model endowed with deep thinking and online search capabilities.

這也是我們對這個模型如何運作有一些見解的地方，他們將其描述為一個具備深度思考和線上搜尋能力的統一多模態生成模型。

So again, just to kind of break that down and simplify it a bit, it's doing the Nano Banana thing in that, you know, it can understand multiple different inputs, be it text or images.

所以再次強調，為了簡化並稍微分解一下，它就像 Nano Banana 那樣，你知道的，它可以理解多種不同的輸入，無論是文字還是圖像。

It can then think about what you're trying to accomplish via your text prompt and then any like holes in its knowledge, it can go online and look those details up.

它就可以思考你透過文字提示想完成什麼，然後它知識上的任何不足之處，都可以上網去查這些細節。

But the crazy thing with see Dance 2.0 is that that architecture extends out to video generation.

但 Sea Dance 2.0 厲害的地方在於，這個架構延伸到了 video generation。

And I can that I've seen it in my own generations.

而且我在自己生成的內容裡也看過了。

It'll report during the process like, hey, I'm looking for this thing that you asked for.

它會在過程中回報，像是「嘿，我正在找你要求的這個東西」。

So essentially what we have in terms of a pipeline is that what like C 2.0 will begin by, you know, reading your prompt, analyzing your input image, video or audio or whatever, and

所以基本上，我們在 pipeline 上的做法是，Sea Dance 2.0 會先，你知道的，讀取你的提示，分析你的輸入圖片、影片或音訊等等，然後

then it kicks over to an LLM planning session where you know, maybe acting as a sub agent, whatever it doesn't understand, it can then kick over to a web search to look for references.

接著會進入一個 LLM 規劃階段，在那裡，也許它會扮演一個子代理，任何它不理解的東西，都可以轉交給網路搜尋來尋找參考。

And again, I'm not going like benchmark bro here or anything, but

而且再次強調，我不是要來 benchmark pro 什麼的，但是

if you look at the C2.0 model card, especially under the video capabilities, it does have a number of impressive features like video reasoning, video perception and motion understanding, long video understanding.

如果你看看 Sea Dance 2.0 的 model card，特別是在 video capabilities 下面，它確實有一些令人印象深刻的功能，像是 video reasoning、video perception 和 motion understanding、long video understanding。

And here's the interesting one, is that this video cut, which allows it to replay relevant segments at a higher frames per second with an intuitive and effective mechanism for Video understanding.

而這個有趣的點是，這個影片剪輯功能，能讓它以更高的影格率重播相關片段，透過直觀且有效的方式來理解影片。

So basically it can replay what it sees at a higher frame rate to understand motion and action.

所以基本上，它可以重播它看到的畫面，並以更高的影格率來理解動作和行動。

I will note this is the part that I'm inferring from the Seed 2.0 model card and the Seed Dream 5.0 research paper actually got that one right.

我要說明一下，這是我從 Sea Dance 2.0 的 model card 和 Sea Dance 5.0 的研究論文推斷出來的，論文裡寫得還蠻準確的。

That essentially the way that this is working is that the seed 2.0 like video analysis part is probably taking place in the narrative, planning and storyboard layer of all of this.

基本上，這個運作方式是，Sea Dance 2.0 的影片分析部分，很可能發生在整個系統的敘事規劃和故事板層級。

So it's actually all happening before we even hit generation.

所以這一切都發生在我們開始生成之前。

So yeah, that is a lot of LLM power and agent work that is all happening under the hood.

所以，這確實是很多 LLM 的強大功能和代理工作，都在幕後進行。

And this is the part where we kind of come back to the cost thing that we were talking about at the beginning.

而這也是我們一開始談到的成本問題。

The only way that ByteDance can run this much intelligence on every video generation is because this thing is so dirt cheap since the mini version of seed 2.0 only costs $0.03 per million tokens.

ByteDance 之所以能在每次影片生成時運行這麼多智慧功能，是因為這個東西超級便宜，Sea Dance 2.0 的迷你版本每百萬個 token 只需 0.03 美元。

You know, that is a structural advantage that makes a thinking video model possible.

這是一個結構性的優勢，讓思考型的影片模型得以實現。

Now, given that, I mean, obviously the big elephant in the room here is that Hollywood is, you know, big mad at sea dance right about now.

不過，說到這個，顯然大家最關心的就是好萊塢現在對 Sea Dance 非常生氣。

But I mean, obviously a lot of key points to their argument are based around, well, I mean, the training on films and TV shows.

但顯然，他們論點的許多關鍵點都圍繞著，嗯，訓練資料使用了電影和電視節目。

And look, I'm not saying that that isn't in there.

而且你看，我不是說裡面沒有。

我的意思是 ByteDance 就是 TikTok。

So I mean if you can think about every movie clip you've ever seen on TikTok, sure that's in there.

所以，想想你曾在 TikTok 上看過的每一段電影片段，沒錯，那些都在裡面。

Plus like a billion people dancing and Every other Rand TikTok video you've ever seen.

再加上數十億人在跳舞，還有你曾看過的每一個隨機 TikTok 影片。

I do want to point out that while TikTok does have kind of a staggering amount of video data, it really is still a drop in the bucket compared to YouTube.

我確實想指出，雖然 TikTok 擁有驚人的影片數據量，但與 YouTube 相比，它仍然只是滄海一粟。

YouTube is, I mean, well over 800 million videos, most of which are long form.

YouTube，嗯，有超過八億個影片，其中大部分都是長篇的。

So when it comes to like just sheer amount of video Data, I mean, YouTube is still king of the hill.

所以說到影片資料的數量，YouTube 還是穩坐龍頭。

Now all of that said, all of that video data might soon be less important than you think it is.

不過話說回來，這些影片資料可能很快就沒那麼重要了。

So something we do know is that Cdance is using retrieval augmented generation or rag.

我們知道的是，Cdance 使用了檢索增強生成（rag）。

you know what that's a joke for the late night show is that it's actually decoupling itself from the training set at inference and then conditioning the generation on whatever it has retrieved.

你知道這就像是脫口秀的笑話一樣，它其實是在推理時從訓練資料中分離出來，然後根據檢索到的內容來生成。

So this is a pretty big deal in that this is a video model that can essentially break free of Its training data.

這很重要，因為這是一個影片模型，可以擺脫它的訓練資料。

Go grab something, bring it back and utilize that to generate.

去抓取東西，帶回來並利用它來生成。

Now, admittedly that is somewhat simplified and I do want to note that it doesn't feed back into the training data.

當然，這有點簡化了，而且我確實想指出，它不會回饋到訓練資料中。

That is something that I definitely have heard a number of times before.

Once the generation is done, it is that retrieved context is gone.

一旦生成完成，檢索到的內容就消失了。

But this does lay the groundwork for a future model that may not rely on training data.

但這確實為未來可能不依賴訓練資料的模型打下了基礎。

我們其實已經有點看到這種情況了。

Well, like on the LLM side, Anthropic's Mike Krieger is the Chief Product Officer over there, said that Claude was essentially writing itself at this point.

嗯，就像在 LLM 方面，Anthropic 的 Chief Product Officer Mike Krieger 說，Claude 在這一刻基本上已經是自我編寫了。

And of course Anthropic CEO Dario Amati has talked about recursive self improvement through the tools that a model has access to.

當然，Anthropic 的 CEO Dario Amati 也談到透過模型可使用的工具來實現遞歸式自我改進。

所以這真的沒什麼好大驚小怪的。

That seed, given its full stack of tools or really any other frontier lab could potentially do this for video in the future.

這個種子，有了它完整的工具堆疊，或者說任何其他前沿實驗室，未來都有可能為影片做到這點。

Yeah, a video model that isn't reliant on training data.

對，一個不依賴訓練資料的影片模型。

It just, it just watches stuff and learns from it.

它就是，它就是觀看東西並從中學習。

Now back over to the Hollywood side of things.

You know, obviously last week Disney and Paramount and a few others sent out their various cease and desist orders to ByteDance.

你知道，顯然上週 Disney、Paramount 和其他幾家公司向 ByteDance 發出了各種停止侵權通知。