2026年了，影视飓风是怎么工作的？

Hey guys, I'm Tim!

It's been almost two years since the last time I shared MediaStorm's workflow.

since the last time I shared MediaStorm's workflow.

In the past year or so, we've made over 200 pieces of content, while constantly pushing both creative and technical boundaries.

Like doing these immersive shoots in all kinds of places around the world, and projects like these super long-form livestreams. All of that has brought quite a few challenges to our existing workflow.

And there's one thing that's impossible to ignore, the rise of AI.

Shots like this, back in the day, they took a ton of time to film.

But now, can we generate them in one second?

Not really.

We still need 500 rounds of 20-second draws, to get visuals like that.

And there are all these large models for coding and researching.

Basically, AI now sounds like it can do anything.

And if you don't use it, you'll be obsolete in no time.

But is that actually true?

Our team has done a deep dive into this, so this time, we want to take this chance to walk you through how MediaStorm's workflow has evolved from 2024 to 2026, especially after bringing AI into the mix.

especially after bringing AI into the mix.

So first, let's start with topic selection.

Two years ago, we already made one thing clear: the core starting point of making a video has to be the title and the thumbnail.

Only after those two are locked in will real people actually click on your video.

And that's still a rule we stick to today.

Before this, coming up with ideas was mostly the producers' job.

We'd regularly hold pitch meetings, and once a producer had an idea, they'd submit a matching title-and-thumbnail combo like this.

But you'll notice, coming up with a great title isn't really a producer's core skill.

A good idea can come from anyone.

So we created this Good Idea groupchat in the company.

No matter what role you're in, as long as you drop in a simple title or even just an image, and people are into it, it could become a real video topic.

That shift has actually made a pretty noticeable difference.

For example, our audience retention and overall views have both gone up quite a bit.

And beyond all the idea work we do ourselves, And beyond all the idea work we do ourselves, after all the AI model upgrades over the past two years, we found that AI assistants, while not great at highly specific execution, while not great at highly specific execution, are incredibly good at gathering information.

So in this groupchat, we used Lark's OpenClaw to deploy a bot like this.

Once you pitch a topic, it quickly pulls together relevant info for you and replies right in the chat or generate a thumbnail, so everyone in the group can see the extra material and build off each other's ideas.

It also actively searches for interesting stuff across the internet.

It's kind of like a gacha system.

We didn't lock it to any fixed sources all these topic ideas are picked by the AI itself from the internet.

And after testing it, we did find some pretty interesting ideas that might be worth filming in the future.

So once the topic is locked, let's look at what changed in planning.

Normally, we plan an episode, the producer handles early research, writing, shot listing, and so on, then we move into production with the script in hand at last.

But you'll notice, scripts in traditional production are always text or image based.

So you spend a lot of time communicating in pre-production, what a shot should look like, how it should be filmed, how the VFX should work.

And it's really easy to waste a lot of effort.

But now, with some help from AI, this new production model has genuinely changed a lot.

For example, our old production flow used to go from text to image, then into filming to become a video, and then into editing.

Now at our company, the creative pipeline looks more like this: we start with text, use AI to turn it directly into a video first, then go shoot it, turn that into a video again, and then edit.

Put simply, during pre-production, we first use AI to generate a reference video, and then build around that to execute something real, a video that people will enjoy.

You can even do things like this, when you're not sure how to block the shots, you can generate a few different versions first and test them out, then pick the best one.

It's very intuitive.

One thing we're pretty firm on is that AI happens in pre-production.

Post is still done by people.

And when you compare it to the final cut, the reference is still pretty different.

But right now, inside our company, AI plays more of a "guiding" role.

Once everyone gets that reference video, they have a pretty clear sense of what they're supposed to do.

It removes a lot of information gaps and cuts down communication costs.

Now, you might be thinking, "Come on, Tim all this AI stuff still sounds pretty abstract.

Got anything practical I can start using right now?"

Alright, now let us show you something we partnered with Lark and built right away.

A teleprompter.

There are all kinds of teleprompters out there.

Some charge you for everything, some are packed with ads.

So we thought, could we build one with Lark that's free, actually good, and uses active AI voice recognition to auto-scroll for you?

So take a look.

This is the system we built with Lark Miaoda.

This is the system we built with Lark Miaoda.

All you do is paste in your script, tap teleprompter mode, and it can scroll with incredible accuracy based on your speaking pace.

We also included reference speaking speeds from different creators, so you can follow their pacing to help guide yourself and speak more naturally.

And more importantly, this teleprompter is completely free.

So if you're interested, just search "Miaoda" in Lark, and you can get access to the teleprompter.

We put real effort into this with Lark, We put real effort into this with Lark, and I think it's worth sharing.

Hope you like it.

Alright, now that we've covered the pre-production side, it's about time to move into production.

And now that we have all these AI-generated visual references, And now that we have all these AI-generated visual references, the entire pace of shooting is much faster than it used to be.

But if we're talking about the actual production process, honestly, there's not that much to say.

It's still setting up cameras, lighting the scene, then hitting record and rolling.

Our backgrounds are all real too, we shoot in real locations, hasn't changed much in the past years.

If you want to ask what's changed the most, it's probably our gear management.

As the team got bigger, the number of projects went up a lot too.

You've probably noticed, we're basically posting daily.

And to keep up with that many shoots and production needs, we've bought a lot more gear.

When we counted it all up, just cameras and lenses alone, we now have this many.

So to manage it all properly, we built an automated check-in / check-out system.

It's kind of like the package lockers, every piece of gear has its own dedicated QR code.

Just scan it with a barcode gun, and it's checked out.

And if you're into gear, we also pulled usage stats for all our equipment, not sure how useful that is to you.

You would notice, the gear we use the most at MediaStorm right now is actually these $1,500 to $3,000 mirrorless bodies.

But the high-end gear that costs tens of thousands mostly just sits in the cabinet collecting dust.

Because with those big dudes, like the ALEXA 35 XTREME, or RED cameras, if you want to actually get them up and running, you have to bolt on a whole pile of accessories.

So it goes from this, to this.

And it gets way heavier.

And right now, weight is probably our biggest pain point.

I used to think once we bought them, they'd pay for themselves right away.

But looking at it now, probably not.

But there were also a few things that surprised me.

At first, they felt super niche.

Didn't seem like they'd ever pay off.

Didn't seem like they'd ever pay off.

But they turned out to be pretty worth it.

Like this whole stack of pro underwater housings.

We bought these housings to handle all kinds of underwater shoots.

And honestly, they're not cheap.

But they really did help us capture a lot of rare footage.

And many clients are willing to pay for this, And many clients are willing to pay for this, because high-quality footage like this is still rare.

And this is exactly the kind of thing AI still can't replace.

Like the great white shark footage.

This is AI-generated.

This is real footage.

The difference is pretty obvious.

It is not a story.

But when you see a great white shark, it's a complete experience.

So for now, AI still can't replace something like that.

And if you really want to say what MediaStorm innovated on in gear, it's probably this underwater audio system.

it's probably this underwater audio system.

You can hear it now.

I'm kind of nervous right now.

It's coming.

It's right above me.

So huge.

Way bigger than I expected.

I thought it'd be smaller.

Oh my god.

That shaky voice from me underwater, and the footage of great white right in front of me, that kind of atmosphere really hits.

And here's this underwater audio setup for you.

It's actually just a DJI Mic 3, plus a small 3D-printed mount.

You fit it inside a full-face dive mask, wear it like this, and you can talk freely underwater.

The setup is simple, but almost nobody in the industry seems to be doing this.

So if you ever need underwater audio, hopefully some of the things we've put together here can be useful to you.

Okay.

There haven't been that many changes in production itself.

So now let's move on to the part that's changed the most, post-production.

Once shooting is done, all the footage gets sent over to our data engineer for backup and transcoding.

Which means card dumping.

In 2025, MediaStorm copied about 1,000,000GB of footage.

MediaStorm copied about 1,000,000GB of footage.

That's 1 petabyte.

This, for example, is a pretty standard 256GB SD card.

If you do the math, that much data would fill about 4,000 cards like this.

Stacked up, that'd be around 8 meters tall.

But that's just for perspective.

Obviously, we're not storing data on SD cards.

So right now, all of our data lives here, in a place we've never shown before.

in a place we've never shown before.

This is Building One's data center.

We ran a 24-core fiber line and linked it to every office building we use.

That gives us two-way data transfer.

Everything copied by the data engineer gets sent through that fiber line into this wall of storage arrays.

And editors can also play back footage smoothly right inside their editing software.

Right now, this rack holds 1,000TB of SSD storage, plus 320 12TB tape drives.

Altogether, that's close to 4,000TB of storage.

Now one thing to keep in mind, while tape is incredibly cost-effective, let's compare the price between tape and SSD right now.

let's compare the price between tape and SSD right now.

Even hard drives are still way more expensive than tape.

But fundamentally, tape isn't that different from the cassette tapes we used to listen to as kids.

tape isn't that different from the cassette tapes we used to listen to as kids.

If you want to find a specific clip, you have to wait for the tape to physically roll there before it can read it.

So access is much slower than SSDs, and there's no true random access.

You can only read and write in sequence.

Right now, we're working on some optimizations.

So if we ever need to pull footage from tape on the fly, So if we ever need to pull footage from tape on the fly, the system can automatically locate it, then move it onto our SSD storage.

That way, when editors are working, they don't really feel the difference.

Just click it, wait a bit, and then play it like normal.

And for this data center, our media tech team has also planned ahead for the future, our media tech team has also planned ahead for the future, we left room for cable routing, and planning to install local compute here too, so we can train our own AI models for the post team to use.

so we can train our own AI models for the post team to use.

And speaking of post, this is probably the part of our company that's been hit hardest by AI these years.

So to avoid getting swept away by this AI tsunami, we've been learning nonstop, and we've also built out a few AI post workflows we can share with you.

and we've also built out a few AI post workflows we can share with you.

and we've also built out a few AI post workflows we can share with you.

Right now, we split our workflow into two categories, AI-led and AI-assisted.

AI-led and AI-assisted.

The AI-led workflow is only used by our shorts team.

The AI-led workflow is only used by our shorts team.

No other department is allowed to use AI-led production right now.

No other department is allowed to use AI-led production right now.

But for shorts, I do think it's one of the places where AI is worth piloting.

Besides shooting regular narrative shorts, they've also been experimenting with AI short films and have even entered AI film festivals, so they've built up a lot of experience.

Shots like these were all generated with AI.

And if you're interested in AI video creation, we're putting together a course focused specifically on AI for image and video, which should be coming later this year.

We've hit plenty of walls putting it together, but I think it's very useful.

so keep an eye out.

As for most of the content on our main channels, the stuff you're watching right now, we're still making it the old-fashioned way.

AI tools are only there to support the workflow.

AI tools are only there to support the workflow.

For example, this is a voice model we trained ourselves this is a voice model we trained ourselves to generate my voice.

You can listen and compare.

Hey guys, what you're hearing right now is my real voice, recorded with a microphone.

Hey guys, what you're hearing right now is generated by my AI voice model.

Pretty close, right?

Some people might ask, Tim, if your model is already this good, are you just never recording again?

I can promise you, these generated voices are only used internally as timing references.

these generated voices are only used internally as timing references.

Editors can use them to rough out the structure of an episode, Editors can use them to rough out the structure of an episode, and clients can review cuts and confirm notes with this version first.

Once the script is locked, that's when we start recording.

Otherwise, for some episodes, I might have to re-record a dozen times, which is brutal on both my voice and my energy.

which is brutal on both my voice and my energy.

But even though AI voices right now are getting close to the real thing, we'd still much rather use a good microphone.

we'd still much rather use a good microphone.

At least today, real audio quality still has a clear edge.

real audio quality still has a clear edge.

For video generation, we're a bit more aggressive with how we use it.

Right now, we're treating it as a support layer for shooting and VFX.

For example, in this episode of Dreams, shots like these probably could've been done with traditional VFX before, but they'd take forever to render, and eat up weeks of someone's time.

So making a full episode that way wasn't very practical.

So making a full episode that way wasn't very practical.

But now, with video generation models, we can run the entire AI video workflow we can run the entire AI video workflow through a custom-built web tool.

through a custom-built web tool.

And all of these dream sequences were made through that AI-assisted pipeline.

A video like this probably wasn't doable before.

Now it is.

And beyond generating video, I think AI is even more useful in VFX.

It doesn't generate, but it massively helps with VFX.

Like green screen keying.

If you've ever done it, you'll know exactly what I mean.

It's miserable.

But about a month ago, one of my favorite teams, Corridor Digital, released a free open-source AI keying tool called Corridor Key.

released a free open-source AI keying tool called Corridor Key.

released a free open-source AI keying tool called Corridor Key.

You just hit one button like this, You just hit one button like this, and it pulls the whole shot cleanly.

Hair, transparent objects, weird backgrounds, it can key all of it in one click.

Kind of insane.

So if you're dealing with keying headaches right now, So if you're dealing with keying headaches right now, I highly recommend giving that workflow a try.

And at this point, you're probably thinking, if AI generation is already this good, then why are you still keying shots?

Why not just generate a Tim too?

Well first, I think the audience still wants me to be me.

I think the audience still wants me to be me.

Right?

The person sitting here should still be a real human.

The person sitting here should still be a real human.

And we have actually tried some of that.

But the result makes one thing pretty obvious, our audience knows exactly what I look like.

They'll spot the cracks no matter what.

So for now, there's just no real way to use an AI version of me for anything serious.

Because our content is supposed to actually connect with you.

Because our content is supposed to actually connect with you.

And you're a real person too.

If you don't buy it, then all of it means nothing.

So overall, So this is what our current AI-plus-traditional post workflow looks like.

Has AI already completely blown up MediaStorm Has AI already completely blown up MediaStorm the way people in the industry keep saying?

Not yet.

Honestly, I think Seedance is right at the edge of something big.

Honestly, I think Seedance is right at the edge of something big.

But it still hasn't fully kicked that door down.

That said, AI is absolutely changing a lot of how we work, AI is absolutely changing a lot of how we work, and how we think.

Like I said earlier, what we work hard to make are things real people actually enjoy.

what we work hard to make are things real people actually enjoy.

So getting real audience feedback and reviewing it fast, matters a lot to us.

Right now, that work is mainly handled by our operations team.

Right now, that work is mainly handled by our operations team.

On top of that, we've also upgraded our old analytics workflow.

we've also upgraded our old analytics workflow.

Inside the company, we created a dedicated role called the data hub.

we created a dedicated role called the data hub.

Their job is to collect post-release feedback and turn it into something visual.

For example, with this Lark dashboard, we can see things like view count and follower growth at a glance, we can see things like view count and follower growth at a glance, and all of it updates dynamically.

That gives us a much more measurable way to evaluate how a piece of content is performing.

In the past, building a clean dashboard like this In the past, building a clean dashboard like this took a ton of time.

But now, all you have to do is say what you want to analyze, and the multi-dimensional sheet can generate a suitable dashboard for you directly from the data we already have.

AI can even plug directly into more advanced analytics tools AI can even plug directly into more advanced analytics tools to pull out denser insights.

And beyond the usual performance metrics, we've also started looking into deeper questions.

Like, what kind of impact does it have when I'm in the video, versus when I'm not?

Especially in the opening.

And in the end, our analytics team came up with a metric called "Pan Presence."

That one got me.

What it measures is the relationship between how long I'm on screen in the intro and the audience retention.

between how long I'm on screen in the intro and the audience retention.

And if that chart feels a little too abstract, the AI built into the multi-dimensional sheet can now just summarize the conclusion for you directly.

And based on the data, it turns out people still like having me around.

Appreciate that.

But we're also testing whether more people on the team can be recognized by more viewers too, so maybe I can step back a little.

Thank you.

And I think at this point, we have to admit it, these AI tools really have had a huge impact on how we work.

these AI tools really have had a huge impact on how we work.

And people are a little anxious.

But I think the only way to deal with that anxiety is to actually try it, or at least understand it.

AI looks complicated, but it's actually easier than it seems. So inside the company, we've set up an internal AI sharing group where people can show and exchange what they've built.

For example, this food ordering tool called "Hungry Dead" was built by one of our colleagues using AI coding tools.

It's not on the level of a pro.

But our view is simple, start using it first, then keep iterating and improving.

Oh, and Lark also recently updated their aily AI assistant.

It's a really powerful AI agent, and we've already used it to build some useful tools.

Like a director's journal assistant.

We're still testing more use cases right now, and if it holds up, we can share more later.

and if it holds up, we can share more later.

Also, Lark recently opened up their official CLI tool, Also, Lark recently opened up their official CLI tool, which makes it much easier for other AI agents to plug directly into our existing workflow.

So if you're interested, definitely give it a try.

And that's everything for this episode.

I'm curious what you think of our workflow, whether any of this was useful to you.

If any of this helped, that would be great.

Don't forget to hit like, subscribe, and share, it means everything to us.

So, until next time!

So, until next time!