New FREE AI Video Generator & Feature Length AI Films!

Summary

この動画は、無料で使える新しいAIビデオジェネレーター「ハイパー」について紹介しています。Google DeepMindの元社員が開発し、1,900万ドル以上の資金調達を成功させたプラットフォームです。テキストからビデオ、画像からビデオ、そしてビデオのリペインティングなどの機能を提供しています。ユーザーインターフェイスは簡単で、HD版とSD版のビデオジェネレーションのオプションがあります。HD版は約2秒間生成され、SD版は4秒間まで生成できます。また、研究によると、AIが完全に生成する映画が近い将来実現される可能性があるとされています。

Takeaways

• 🌟 新しいAIビデオジェネレーター「ハイパー」が登場し、現在完全無料で利用できます。
• 🤖 ハイパーは、Google DeepMindの元社員であるYia MeowとZuu Wongによって開発されました。
• 💰 ハイパーは1,900万ドル以上の資金調達を行い、AIビデオ競争で大きな存在感を見せています。
• 📝 ハイパーはテキストからビデオ、画像からビデオ、そしてビデオのリペインティング機能を提供しています。
• 🎨 ハイパーのインターフェースはシンプルで、フルHDのビデオジェネレーションや標準定義のオプションなど、多様な機能があります。
• 📹 ハイパーで生成されるHDビデオは約2秒間で、延長方法としてエディターでのスローモーションを提案しています。
• 📚 研究では、GPT-4と画像からテキストのモデルを利用して、詳細な脚本と対応するビジュアルを生成する「ムービーLLM」が提案されています。
• 🎬 ムービーLLMは、映画のプロット、スタイル、キャラクターを生成し、一貫性のあるスタイルでシーンを制作することを目指しています。
• 🔍 ムービーLLMは、1,000以上の映画と60,000以上のトレイラーのデータセットを利用して、より良い映画の概要を生成しています。
• 💬 ミッドジャーニーのCEO、David Holtzは、不審なアカウントからのアクセスが原因で24時間サービス停止が発生したと説明し、Stabilityの従業員からのアクセスを禁止しました。
• 🔥 StabilityのImad Mustakは、チームがスクラッピングを行っていないと主張し、自社のデータセットと拡張機能に満足しています。

Q & A

ハイパーとはどのようなプラットフォームですか？

ハイパーは、テキストからビデオを生成する新しいビデオ生成プラットフォームで、元Google DeepMindのメンバーによって作られました。

ハイパーの開発に携わった人物は誰ですか？

ハイパーは、Yia MeowとZuu Wongという元Google DeepMindのメンバーによって開発されました。

ハイパーはどのように資金調達を行ったのですか？

ハイパーは、総額1900万ドル以上の資金調達を行いました。

ハイパーで提供されている機能は何ですか？

ハイパーではテキストからビデオ、画像からビデオ、そしてビデオのリペイント（再描画）を提供しています。

ハイパーのインターフェイスはどのようになっていますか？

ハイパーのインターフェイスは比較的直截了しており、ライトモードとダークモードの切り替え、プロンプトボックス、ビデオ生成オプションが用意されています。

ハイパーで生成されるHDビデオの長さはどのくらいですか？

ハイパーで生成されるHDビデオは、現在は約2秒程度の短いものとなっています。

ハイパーの標準解像度（SD）モデルとHDモデルの違いは何ですか？

SDモデルは、より長い（最大4秒）ビデオを生成できますが、HDモデルのように動的な結果を提供するわけではありません。

ハイパーのコミュニティフィードで見られる成果物はどのようなものですか？

ハイパーのコミュニティフィードでは、高品質なビデオクリップやアニメーション、スタイリッシュなシーンなどが見られます。

ハイパーのビデオを長くするためにどのような方法がありますか？

ハイパーのビデオを長くするために、非線形エディター（例：Premiere ProやDaVinci Resolve）を使用して、ビデオのスピードを低下させ、オプティカルフローを有効にすることで、ビデオを延長できます。

ムービーLLMとは何ですか？

ムービーLLMは、GPT-4と画像からテキストのモデルを利用して、詳細な脚本と対応するビジュアルを生成する研究プロジェクトです。

ムービーLLMがどのように機能するのですか？

ムービーLLMは、GPT-4で映画のテーマや概要、スタイル、キャラクターを決定し、その後、シーンごとに詳細な脚本を生成します。その後、Stable Diffusionを通して、シーン、キャラクター、場所のスタイルを固定し、一貫性を保ちながらキーフレームを生成します。

ムービーLLMが使用したデータセットは何ですか？

ムービーLLMは、1000以上の映画と60,000以上のトライラーからデータを作成したMovieNetデータセットを使用しています。

Outlines

00:00

🌟 AIビデオジェネレーターの紹介

この段落では、無料で使える新しいAIビデオジェネレーター「Hyper」について紹介されています。Google DeepMindの元社員であるYia MeowとZuu Wongが立ち上げたプラットフォームで、1,900万ドル以上の資金調達を成功させています。テキストからビデオへの変換、画像からビデオへの変換、ビデオのリペイントなど、様々な機能を提供しています。ユーザーインターフェースはシンプルで、標準的なオプションとHDバージョンのビデオジェネレーションが可能です。しかし、HDバージョンのジェネレーションは2秒程度の短い制限があります。

05:00

🎬 AIビデオの品質と特徴

この段落では、HyperのAIビデオジェネレーションの品質と特徴について詳しく説明されています。特に、車の描写やゴーストトレインのような創造的な要素が好評を得ています。また、AIビデオの短さに関連して、非線形エディターを使用してクリップを延長する方法も提案されています。最後に、完全なAI映画の生成に向けた研究と、その過程でのスタイル固定化プロセスについても触れられています。

10:02

📢 Mid JourneyとStabilityの対立

この段落では、Mid JourneyのCEOであるDavid Holtzが、24時間のサービス停止がStabilityの従業員によるボットのような行為によるものであると説明し、Stabilityの全従業員がMid Journeyサービスの使用を禁止したことを発表しています。Stabilityの代表であるNick S. PierreとImad MustakがTwitterで対立し、データセットの品質やモデルの性能について主張し合っています。また、Mid Journeyがビデオモデルをトレーニング中であり、StabilityがStable Diffusion 3をリリースする予定についても言及されています。

Keywords

💡AI视频生成器

AI视频生成器是一种利用人工智能技术自动创建视频内容的工具。在视频中，提到了一个新的、完全免费的AI视频生成平台，它允许用户通过文本提示生成视频。这个工具的出现预示着AI在创意产业中的应用越来越广泛。

💡文本到视频

文本到视频（Text to Video）是一种AI技术，它可以根据用户提供的文本描述自动生成相应的视频内容。这种技术在视频制作和内容创作领域具有革命性的意义，因为它极大地降低了视频制作的门槛。

💡视频重绘

视频重绘（Video Repainting）是指对现有视频内容进行风格上的改变或创新，使其呈现出全新的视觉效果。这种技术在艺术创作和视频编辑中有着广泛的应用。

💡Mid Journey

Mid Journey是一家提供AI创作工具的公司，它的服务在视频中被提及，因为与稳定性（Stability）公司之间发生了一些争议。这个争议涉及到了数据抓取和使用的问题。

💡稳定性（Stability）

稳定性（Stability）在这里指的是一家与Mid Journey发生争议的公司。这个名词在视频中被用来描述两家公司之间的紧张关系。

💡GPT-4

GPT-4是一种先进的自然语言处理模型，它能够理解和生成人类语言。在视频中，GPT-4被用于电影LLM（Movie LLM）项目，以帮助生成电影剧本和视觉内容。

💡风格固定化（Style Immobilization）

风格固定化是一种艺术和技术手段，用于确保在创作过程中保持一致的视觉风格。在视频中，这一概念被用于AI生成的电影中，以确保电影的视觉风格和角色形象保持一致。

💡特征长度电影（Feature-Length Movies）

特征长度电影指的是时长符合电影标准的作品，通常在90分钟以上。在视频中，讨论了AI技术在生成这种长度电影方面的潜力。

💡数据集（Dataset）

数据集是用于训练和测试机器学习模型的数据集合。在视频中，提到了MovieNet数据集，这是一个包含超过1000部电影和60000个预告片的庞大数据集。

💡合成数据（Synthetic Data）

合成数据是通过算法生成的数据，它可以模拟真实世界的数据特征。在视频中，提到了合成数据在AI模型训练中的应用。

💡非线性编辑器（Nonlinear Editor）

非线性编辑器是一种视频编辑软件，它允许用户以非线性的方式编辑视频，即可以在任何时间点进行编辑，不受时间顺序的限制。

Highlights

New free AI video generator introduced.

Hyper developed by Yia Meow and Zuu Wong, ex-Google Deep Mind.

Hyper raised over $19 million.

Offers text to video generation and image to video.

Hyper interface includes light/dark mode and various video generation options.

Standard definition model allows up to 4 seconds of video.

Community feed showcases diverse user-generated content.

Potential for AI-generated feature-length movies discussed.

Movie LLM research paper uses GPT-4 and text-to-image models.

Style immobilization process for consistent movie generation.

MovieNet dataset used for training in the paper.

Mid Journey CEO addresses 24-hour outage and bot-like behavior.

Ban on Stability AI employees using Mid Journey.

Mid Journey video model in training, reported to be good.

Stable Diffusion 3 release by Stability AI imminent.

Transcripts

hey everyone today I have a brand new

totally free AI video generator it's

pretty impressive and I've got some

interesting information on who's behind

it and yes I should say totally free for

now you guys have been around the block

long enough you know how this works plus

we're going to take a look at the latest

research on generating fulllength AI

films yep it's coming all that plus

drama at Mid Journey yeah they've got

bad blood with stability. a this one is

pretty wild it's actually rumored to be

the plotline for the second season of

Netflix's beef okay lots to cover let's

dive in first up is hyper. a a new video

generation platform that you can start

using right now totally for free this

one is brought To Us by Yia meow and zuu

Wong two former Google Deep Mind alums

who have teamed up to form this platform

raising over $19 million so it does look

like hyper is set to be quite a

contender in the AI video race hyper

offers text to video uh very ni nice

clean smooth movement here we're going

to take a deeper look at that but

clearly I mean obviously it does these

animated Styles very well it also does

image to video in this example uh they

actually showcase that you can generate

images on the platform and then animate

that I have not seen that on the

platform yet but we are going to take a

look at some image to video examples and

also somewhat surprisingly they offer

video repainting or you know video in

painting essentially they actually had

this kind of neat demo of repainting on

the site with you know somebody pouring

the smoothie mixture into a bowl and

then repainting it to be like this

watercolored uh you know koi fish thing

uh yeah very creative like it a lot

getting started the interface is pretty

straightforward there is a light dark

mode down here by the way uh change it

to dark because once again only

Psychopaths use light mode your prompt

box is down below and you have a number

of different options in terms of how you

want to generate your video uh either in

a full HD version animate your image

which is basically image referencing

repainting your video and then two more

options for creating with text prompts

or image referencing but this time in

standard definition there is a reason

why we'll get to that in one second

finally at the end of the options is

extend your video uh this obviously has

not been enabled yet but will be coming

soon diving in in my last video I did

kind of a quick sketch music video of

some creepy dolls working in a factory

for that I ended up utilizing Pika so I

decided to take the same prompt and run

it through hyper just to see what the

results would look like so pretty much

the exact same prompt which is stop

motion facel dolls working in a factory

Gothic dark uh yielded these results

which looks really good so to note

Generations in this HD format tend to be

around 2 seconds I know it's super short

I do have a bit of a workaround for that

we'll get to that in just one minute

once you're in the generation and

creation area if you want to swap out

modes you can simply come up to this

creation mode up here and then you know

you have all of your various templates

once again so taking that same prompt in

the standard definition definition model

you'll note that we now have additional

templates uh up here that we can add in

basically what this is doing is it's

just going to add keywords into your

prompt so uh say we're going with this

in steampunk uh by hitting that we now

have weighted steampunk style uh

mechanisms fantasy gear decoration Etc

we also have additional options here uh

for example you can lock the seed number

uh there is controls for motion levels

let's just crank that all the way up and

then the duration on the standard

definition version you can take your

Generations up to 4 seconds running that

but removing the steampunk prompts uh

yields this which yeah that looks pretty

good I would put it pretty much on par

with the Pika outputs that I got in the

last video so these are definitely two

different models image to video also

looks really good I took this Dune

inspired like heavily Dune inspired

image that I generated up in mid Journey

mostly because I have Dune fever I have

not seen it yet hopefully next week

anyhow running that into hyper and then

adding the prompt man looks to camera

intense gaze and running that in the HD

mode yielded this which yes while short

does look very good now taking that same

image and running it through the SD

model with the exact same prompt yielded

this the results from the SD version are

just a little bit more on the I know

kind of static and Bland side uh

additionally Timothy shalom's stunt

double here has kind of an animated look

to it his eyes are kind of doing the

Pennywise thing as well uh and just you

know overall the HD version is

definitely far superior to this that

said I do think that you can still get

some really cool results out of the SD

model uh for example this is a prompt of

just uh a wizard standing in an

enchanted forest and I use the gibli

preset and yeah this looks really really

nice hopping over to the Community Feed

to see the breath of what hyper is

capable of uh I ran across this output

which looks really really good um yeah I

mean I would hesitantly call this

somewhere in the neighborhood of Sora

quality although you you know granted

much shorter something that I noticed is

that the cars are all maintaining

consistency like they're they're not

morphing all over the place and there

aren't a few of them like driving

backwards although I do have a bone to

pick that there are what like five

Outward Bound lanes and one inward bound

Lane there's one thing Sim City taught

me is that this is going to end in

disaster sticking with the imaginative

side this ghost train in the sky looks

really really good this interior zoom

out also looks really good although

obviously you know Limited in time it's

it's also not like morphing out too much

everything is staying consistent and

stable this extreme closeup also caught

my eye yeah I mean pun not intended uh

but yeah this is actually text video

it's not image to video so that's what

I'm saying about it approaching sore

level quality although just you know not

as long few more quick ones yeah walk

cycle here looks really really good uh

all six legs look like they are actually

present and not just like sort of

floating along a surface another walk

cycle or I should say step cycle uh in

sort of an animated style it's really

solid and finally cribbing one of the

famous Sora text prompts fly over of a

California Gold Rush town uh we end up

with this which yeah I mean this looks

really really good now I should say it's

still AI video and there is still

weirdness abound uh taking our old

friend the man in the blue business suit

walking down a busy City street gets us

this uh where our guy is like

moonwalking down the center lane of a

city street definitely not safe uh

additionally we have this car over here

that's actually driving backwards but

you know maybe they just made a wrong

turn down a one-way street but again at

the cost of free I mean roll to your

heart's content until you find something

that you like overall I think the

results out of hyper are super

impressive yeah there is that 2cond

limitation on the HD versions uh but

that is a bridge that will be crossed at

one point or another in the meantime uh

if you want a quick hack in terms of

extending your shots you can always

bring your clip into a nonlinear editor

like Premiere that we have here or Da

Vinci resolve and extend it out in

Premiere what you'll want to do is uh

just right click on it come to speed

duration uh take the speed down to 50%

and then just make sure that Optical

flow is turned on here Da Vinci also has

something pretty similar uh but yeah as

you can see the results are pretty good

A little slowed down but I mean not bad

ultimately to get the best results you

should probably use something like topaz

video Ai and listen I know it is super

expensive but it is also kind of the

best in class for for this particular

job and that pretty much is the

smoothest and best result that you can

possibly get and again I know the 2C

thing is a bit of a bummer but again

that is just temporary and even now I

think you can craft together a pretty

solid narrative by using a combination

of 2cond HD shots extending them out a

bit and 4 second SD shots uh to vary out

your shot lengths anyhow moving on from

like two second shots to feature length

movies uh we have movie llm enhancing

long video understanding with AI

generated movies now this one is in the

paper research Arena but it does set the

stage for what will be eventually coming

which is you know fully generated AI

movies from a prompt we actually have

seen something kind of similar to this

in a previous video namely LTX studio uh

if you missed that video it is linked

Down Below movie llm leverages The Power

of GPT 4 and text to image models to

generate detailed scripts and

corresponding visuals the work path of

movie llm is the movie plot generation a

style immobilization process and video

instruction by data generation movie llm

leans on GPT 4 for sort of the initial

breakdown of the film um with GPT coming

up with the theme the overview movie

style frame level description and the

characters of the film from there GPT 4

takes that output and generates up Epoch

chapters essentially scenes for the most

part uh at which point it takes it

through another process where it will

generate out you know chapters for each

scene detailing both the characters the

actions that are happening and the

location uh from there it actually takes

that output and then begins generating

dialogue for each of those scenes the

real key to all of this is in the style

immobilization process in which uh

essentially keywords are then extracted

from each of those chapter epochs the

characters and the plot summary uh run

through stable diffusion to generate up

the various you know scenes characters

and locations uh but then taken through

the uh textural inversion process and

the immobilized style embedding in order

to essentially lock the look of the

entire film into one consistent or what

multiple consistent characters and

multiple consistent locations from there

it's taken back out through generation

guiding and through another stable

diffusion process to generate out key

frames for each one of the scenes now

the paper did not provide any video

examples but we can see key frame

comparisons uh namely uh the movie llm

output is down here where you know it

does it looks like it's very consistent

in terms of style and in character

whereas in the other two examples uh

this custom diffusion one in particular

is a little bit all over the place in

terms of its overall style and look just

zooming in here for a minute because the

custom diffusion model actually did make

me laugh a little bit like in the second

example where he's supposed to be a

blacksmith he just kind of looks like a

drummer for like an 80s rock band

interestingly in the paper they do note

that there isn't a lot of resources out

there in terms of training models for

extremely long like featurelength movies

uh but they ended up using the data from

the movie net data set which apparently

is a massive data set of over 1,000

movies and 60,000 trailers now to note

the movies and trailers aren't actually

within this data set it's just the data

from those films and trailers by

utilizing the movet data set movie llm

is able to basically come up with better

synopses for its own films weirdly

flexed with an example by using the film

splice I've seen splice and that's a

pretty odd choice to go with for an

example if you haven't seen it I pretty

much say you know you can take a pass on

that one uh what you should check out is

natali's other movie Cube that movie is

really good ring out with some pretty

odd mid Journey news uh yesterday's

office hours started off on a pretty

weird note mid Journey CEO David Holtz

explained that the 24-hour outage they

had on Saturday was caused by botlike

behavior from Paid accounts he then said

that this originated from stability. a

employees basically scraping for images

and text prompts at which point he

announced that he was effectively

banning all stability. employees from

using the mid-journey service Nick s

Pierre took to Twitter as this was

happening and offered stability. imod

mustak the opportunity to comment imad's

response was uh what mod shortly

followed up with and I'm just quoting

directly here it's Twitter there are

grammatical and spelling errors glore

but you know it again it's Twitter uh

very confusing how two accounts would do

this team also hasn't been scraping as

we have been using synthetic and other

data given sd3 outperforms all other

models anyway I am a big mid journey and

David fan which is why I back them at

the start with the grant to pay for the

beta on E we go IM mod followed up from

there with if anyone did do this on team

have asked we'll dig also happy if mid

Journey reaches out direct it's not

great but obviously not a DDOS attack

but unintentional certainly not

instructed by us/ stability though

really happy with our data set and the

augmentation we have on that a little

while later David did respond to Ahad

saying sent you some information to help

with your internal investigation

definitely interesting I do hope the

boys can resolve this issue I mean we

have enough going on with Elon and Sam

we do not need any more drama meanwhile

mid journey is still training their

video model it is reported to be quite

good and stability will be releasing

stable diffusion 3 into the wild uh

pretty much at any moment so that wraps

up another crazy week in Creative AI as

as always I thank you for watching my

name is

Tim