跟读练习: SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer, [ICLR 2026, Oral] - 通过视频学习英语口语

选择练习模式

每种模式都会打开全屏练习室。

125 句

如果句子过短或过长，请点击 Edit 进行调整。

Thank you, Luke, for the amazing opening.

So good morning and thanks for having me.

My name is Jun Song.

I'm from the University of Hong Kong and NVIDIA.

It's really my great honor to present this work, SANA Video, for our SANA team.

So currently, generative AI is evolving rapidly with model sizes growing larger and larger.

At At the same time, AVIDIA GPUs are becoming increasingly powerful.

We can foresee that the future visual generative models will demand even more computational resources.

So actually, generating images and videos are significantly more computationally demanding than processing just text.

For example, the inference for Flux or CogVideoX exceed the Lama 3,

and an order of magnitude higher than the VIT models.

So they are restricted to cloud deployment and remain impractical for consumer GPUs or age devices.

So actually, efficiency is a forever topic.

So far, we have explored several directions for efficient video generation.

First, in Sanaa, we tried to reduce input tokens by using our self -developed deep compression autoencoder, DCAE.

We also try to use linear tension to replace Softmax 1, enhance the architectural efficiency of the model.

And we also already explored inference step distillation in Sanaa's spring project.

The fewer the inference steps, the faster the speed.

So the last part is what we are going to talk about today, the hybrid AR plus diffusion for long video generation.

Actually, we can generate video block by block causally and diffuse within the block.

So video generation requires much more computation than images.

SANA can generate one image within just one second.

However, generating a five -second video with one model actually takes nearly 30 minutes.

So to boost the video generation efficiency, we focus on three key technical areas, the high ratio VAEs and leaning attention mechanisms,

especially optimized for video tasks.

Crucially, the block KV cache based on leaning attention enables the generation of much longer videos.

So to get a video generation model, actually, we have two ways to go to try to choose.

First, we can just train from scratch.

But only when we have hundreds or thousands of GPUs,

like H100 GPUs, and it will take you almost several months to train one model.

And it's, of course, not affordable.

Or actually, we can start from a semantic aligned text -to -image model.

Here we start from our SANA 1 .5, which is a 1 .6 billion parameter text -to -image model trained on our DCAE.

Actually, since our video model is based on image model, so the linear tension DIT actually is almost the same as our SANA,

but we're introducing two modules for temporal modeling, and we will talk about them later.

Finally, on the left at the top, we implement block causal linear tension module to achieve constant kv cache memory.

So for any length video generation, So first, let's talk about linear attention in video generation model.

Actually, in video tasks, linear attention provides even more significant acceleration than in images,

as the video length and resolution are much larger.

Our results show that for long videos, SANA video is 1 .5 times faster than the one model.

And additionally, at high resolution, our linear attention is four times faster than the flash attention.

And these improvements make a long form, high resolution video generation much more practical.

However, we find that the linear attention map is far different from the softmax attention.

As you can see in this figure, the linear attention is super dense, while the softmax attention is rather sparse.

And this demonstrates that the local contacts captured by the linear tension is actually insufficient.

To enhance the temporal modeling, we introduced two key improvements.

First is the 3D rope for the linear tension and temporal convolutions.

And both are highlighted in the orange box.

While 3D rope doesn't solve the tension map sparsity problem entirely, it do help improve attention low quality.

Combined with our spatial temporal mix FFN, it significantly boosts the overall performance.

Here we have the training loss.

The green line stands for the 3D rope and the additional temporal convolution.

And to further support long video generation, we leverage the unique properties of linear attention to implement an efficient block KV cache.

So let's quickly review how this works.

In linear attention, we call it the attention map, or the state.

Actually, it has a fixed shape of D times D, regardless of the input sequence length.

So when processing the first block here, we obtain the state, S1.

And for the second block, the state is updated additably,

meaning that we only need to add the current block's contribution to the previous cache.

And generally for any block I,

actually the current state is updated as adding the current block contribution the KV here to the previous cache to ask.

So now let's see how we enable long video generation with block of causal linear attention.

For the first block here, the green one, we compute the state as one, shown in the green box.

After we compute first block the state, we cache it, it's shown in the red box.

Then we calculate the KV for the next block and add it to the previous one.

Here we take the cached state from block one in the red box and updates it into the block two, which is in the green color.

And the subsequent process follows the same logic.

The new state is cached, and once the next block competition is finished, they are summed together. And so on.

So, token in Valenian attention will become larger, more and more as the video length become larger and larger.

While our causal linear tension maintains a constant KV cache potentially make minute length video generation possible.

So here in conclusion, we can have two advantages of block linear tension.

First, we have constant memory usage regardless of the input sequence length and second we have the global history context.

So one last question is how to achieve causal FFN with temporal conf.

So let's take two block situation as example.

During causal training the block one cannot foresee block two so we pad a zero token here between the two blocks.

However block two must see block one's token so that it's correct the causal modeling.

So here we catch the last token from block one as the first token in block two.

So finally the two blocks are separated by the zero padding token

which is subsequently removed to to avoid affecting the final computation.

And the block two used the last token from block one making the the causal modeling available.

So here's some results, compared to the current state -of -the -art video models, Sonar Video performs competitively on VBench,

and it takes only 18 seconds to generate a 1K resolution video on a single H100 GPU.

Here we show some examples.

On the left we have text to video examples, And on the right, we show the image to video examples.

And using SANA video as a base model, it can be applied to tasks like physical AI and real -world models.

Actually, by fine -tuning SANA video on a small data site with just a few safety steps,

we enable it to generate videos for robotic manipulation and gameplay play.

We also combine with the post -training tech in long lab, which is a three long test long method.

And this method is also accurate poster and is also from our group.

We also implement the DMD step distillation to further accelerate our model.

And finally, Sonar Video you can also generate minute -length video within just 36 seconds,

achieving 27 FPS real -time generation.

So, finally, there is one more thing.

Based on linear attention, we have much more potential applications to discover.

We believe the efficiency of linear attention

and deep compression auto encoder will facilitate streaming video generation

and also the development of long form like minute -length or even hour -length, real -world models.

And we are working on this project and we'll see if we can keep pushing the efficiency boundary.

And all of our projects, including the code and checkpoints,

100

are all open -source to encourage further exploration in efficient generative AI system.

101

So that's all for my talk today.

102

And we will have a poster session right after this aural, maybe in the afternoon.

103

Welcome everyone to discuss more.

104

Thanks so much.

105

Thank you.

106

Time for a quick question, if any.

107

Yeah, please go to the mic.

108

Thank you for your speech, and I have a question.

109

I think it's something similar to the auto -aggressive video diffusion models like cause of it

110

and something similar to the block causal linear attention you introduced.

111

So I wanted to ask what are the key differences and if we are thinking about infinite long video generation,

112

which one do you think would be better choice?

113

Thank you.

114

Thank you for your question.

115

Actually the cause V something like that, they actually perform on soft max extension.

116

So as I talked before, the soft max extension has an open square computation complexity.

117

So if you want to generate mini lens or even hour lens video, it's not possible to do it on soft max extension.

118

So if you want to generate mini lens video, so you have to keep the computation within like a sliding window or something like that.

119

So the main difference here is our linear tension has the global context.

120

It has memory from the first second to the final second.

121

So if you want to do some project or task like world model,

122

it has to be maintaining overall all the cache or the memory.

123

So I think that's the main advantages of getting attention.

124

Yeah, thank you.

125

Thank you.

下载应用

流利口语所需的一切

AI 发音评分逐句打分

音标练习掌握每个音

词汇学习建立你的词库

词汇游戏边玩边学

扫码

Download on theApp Store Get it onGoogle Play

✨ 推荐视频

33 句

Enhancing AV Simulation with Generative AI - DRIVE Labs Ep. 36

Daily ConversationUniversity StudentsWorking Adults

321 句

Mixture of Experts Explained Visually: How Trillion-Parameter Models Actually Work

Daily ConversationUniversity Students

747 句

本课的词汇与口语要点

这节口语课以视频“SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer,”为素材。视频中反复出现的词有：block, attention, model, linear, generation。这段视频共有 125 个句子、1634 个单词可供跟读。讲话部分时长为 12:46。说话人语速平稳，每分钟约 128 个词，很适合跟读。只有 80% 的单词属于英语最常用的 3,000 词，词汇难度较高。

视频中的重点词汇

视频中 15 个值得学习的单词，附发音和释义：

单词	发音	释义
linear 形容词	/ˈlɪn.i.əɹ/	线性
cache 名词	/kæʃ/	緩存 /缓存, 快取
tension 名词	/ˈtɛnʃən/	緊張 /紧张, 不安
token 名词	/ˈtoʊkən/	代幣 /代币, 籌碼 /筹码
temporal 形容词	/ˈtɛm.pə.ɹəl/	世俗的
efficiency 名词	/ɪˈfɪʃ.ən.si/	效率, 效能
efficient 形容词	/ɪˈfɪʃənt/	高效率
introduce 动词	/ˌɪntɹəˈdus/	介紹 /介绍, 紹介 /绍介
input 名词	/ˈɪnpʊt/	輸入 /输入
implement 动词	/ˈɪmpləmənt/	實施 /实施, 執行 /执行
lens 名词	/lɛnz/	透鏡 /透镜, 鏡片 /镜片
rope 名词	/ɹoʊp/	繩子 /绳子, 繩索 /绳索
sequence 名词	/ˈsiː.kwəns/	序列, 順序 /顺序
contribution 名词	/ˌkɑntɹɪˈbjuʃən/	貢獻 /贡献
explore 动词	/ɪkˈsploɹ/	探索, 探險 /探险

视频中的语法

说话人最常用的结构，并附上视频中的原话：

结构	视频中的用法
被动语态 be + 过去分词 — 强调发生了什么，而不是谁做的	are restricted · is based · are highlighted
定语从句 who / which + 从句 — 补充说明人或事物	two, which is · lab, which is

需要注意的发音

说话人用了 4 次缩略和弱读形式，例如 doesn't, I'm, we'll。请按听到的简短形式来说。

“sh” 和 “zh” 音: cache /kæʃ/, tension /ˈtɛnʃən/, efficiency /ɪˈfɪʃ.ən.si/, extension /ɪkˈstɛnʃən/, efficient /ɪˈfɪʃənt/
长单词——注意重音位置: efficiency /ɪˈfɪʃ.ən.si/, contribution /ˌkɑntɹɪˈbjuʃən/, automobile /ˈɔː.tə.məˌbil/, potentially /pəˈtɛnʃ(ə)li/, mechanism /ˈmɛk.ə.nɪ.zəm/

中文母语者容易读错的音：

/v/ — 不要读成 /w/，上齿轻触下唇: solve /sɒlv/, discover /dɪˈskʌvə/, aggressive /əˈɡɹɛs.ɪv/, leverage /ˈlɛv.(ə.)ɹɪd͡ʒ/, evolve /ɪˈvɑlv/
词尾辅音 — 要读清楚，后面不要加元音: cache /kæʃ/, generate /ˈd͡ʒɛn.ə.ɹeɪt/, efficient /ɪˈfɪʃənt/, introduce /ˌɪntɹəˈdus/, input /ˈɪnpʊt/
/r/ 和 /l/ 的区别: linear /ˈlɪn.i.əɹ/, temporal /ˈtɛm.pə.ɹəl/, regardless /ɹɪˈɡɑɹd.lɪs/, explore /ɪkˈsploɹ/, practical /ˈpɹæk.tɪ.kəl/

如何用这段视频练习

先完整听一遍视频，不要开口，记下不认识的单词。
用正常速度逐句跟读，每句重复到你的节奏与说话人一致为止。
录下自己的声音并与原声对比，特别注意 linear, cache, tension 这类单词。

什么是跟读法？

跟读法 (Shadowing) 是一种有科学依据的语言学习技巧，最初开发用于专业口译员的培训，并由多语言者Alexander Arguelles博士普及。这个方法简单而强大：您在听英语母语原声的同时立即大声重复——就像是一个延迟1-2秒紧跟说话者的影子。与被动听力或语法练习不同，跟读法强迫您的大脑和口腔肌肉同时处理并模仿真实的讲话模式。研究表明它能显着提高发音准确性，语调，节奏，连读，听力理解和口语流利度——使其成为雅思口语备考和真实英语交流最有效的方法之一。

影子跟读法: 阅读完整分步指南 →

跟读练习: SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer, [ICLR 2026, Oral] - 通过视频学习英语口语

选择练习模式

下载应用

Enhancing AV Simulation with Generative AI - DRIVE Labs Ep. 36

Mixture of Experts Explained Visually: How Trillion-Parameter Models Actually Work

MIT Introduction to Deep Learning | 6.S191

Why Memory Prices Are Crazy High

Controlled Multi-modal Image Generation for Plant Growth Modeling

Trust the Process: How to Design Teams That Actually Work

Neural Radiance Fields in 2 Minutes

How I Learned Procedural Generation

The future unfolds! How to optimize Android apps for adaptive layouts

LangChain in 5 Minutes (Explained Clearly)

What AI can (and can't) currently do in a real lab | Jennifer Doudna

NVIDIA GTC 2025 - Keynote Highlights by NVIDIA CEO Jensen Huang

本课的词汇与口语要点

视频中的重点词汇

视频中的语法

需要注意的发音

如何用这段视频练习

什么是跟读法？