Keling AI in 4K Couldn't Revive the Brutal Underclass Story

"Gimme that 480P**"**

Never talked about Keling AI before. Keling AI gives me the vibe of that one kid from my hometown — you know, the small-town tough guy who rolls with the rough crowd. You only have a vague impression of him, know he carries some clout in certain circles, but you'd never think to hit him up for everyday stuff.

Be like: "Yeah, I don't do social like Dreamina, I don't game the leaderboards like so-and-so, I'm just a triple-specialized video foundation model focused on serving professional creators making professional videos." No wonder you average users don't have me on your radar.

They've been marketing themselves on physics simulation, cinematic quality, long-form video — basically implying that sure, other models are fine for messing around, but if you actually want to make a proper film, you come to Keling AI. Recently they dropped native 4K generation, which feels like they're doubling down on monopolizing commercial AI video production.

Keling AI had another little moment recently — apparently foreigners are obsessed with using it to make those "Korean girl at a baseball game" videos, with realism that's supposedly surpassing Western rivals.

Checked their homepage and — oh, they just released a template.

Anyway, we decided to put it to the test and see whether our guy's got confidence or just delusions. Speaking of Kuaishou, one of the internet's great regrets is that those wild, unhinged videos from the "cruel underbelly" era — born in a time before HD smartphones and 5G — are just too low-res to properly appreciate.

480p firecracker-to-the-crotch footage where you can barely hear the bang, pixelated and faded videos of that guy eating... you know what, where you can't even tell what's real. Honestly, who wouldn't want to be there in person, to feel the raw power of those cruel underbelly tales?

So I decided to pair Keling AI's new native 4K output with 1:1 recreations of classic works from Knife Bro, Tiger Bro, and company — taking us all back to those unrestrained days.

First up, the classic of classics: "Fierce Tiger Crosses the River."

Specifically: Tiger Bro with just his head sticking out, wrapped in a burlap sack, his disciple gives him a hard kick, and Tiger Bro launches into the river, never to be seen again. Legend has it he was rescued downstream by an auntie washing vegetables.

I uploaded a clip of Tiger Bro's video and audio, established his likeness, cloned his voice, and crafted prompts based on the original video.

Oh right — to honor how Knife Bro and Tiger Bro back in the day only knew to beg for hearts, not making a single dime, I specifically set the scene in a Kuaishou livestream room, letting digital audience members send gifts to make up for their unpaid native careers.

Final product:

At the start, when the disciple is packing up Tiger Bro, his hands are doing way too much — like a Douyin influencer who suddenly broke into finger choreography.

Later when walking, it's also kinda stiff, feels like an Xpeng robot wearing human skin.

As for the sack hitting the water, the physics are actually pretty solid — the waves, the splash sounds, the ripples all met expectations.

But the most serious problem: didn't I say this was a Kuaishou livestream? Why'd they just send me back the original video?

So I tried Dreamina to see if the livestream shell issue was universal.

Turns out they can generate it. Though the text it generates is basically gibberish.

Beyond that, with the same audio uploaded, Dreamina's voice cloning is just better than Keling AI's.

Downside is Dreamina likes to add its own creative sound effects based on independent thinking — like this fart noise, absolutely no idea where that came from.

Second test: "Donald Duck Hangs Himself."

As the name suggests, Tiger Bro hangs his new disciple Donald Duck off a bridge, kicks away the stool, Donald Duck starts struggling, Tiger Bro screams "Double tap!"

Final product:

This round Tiger Bro's speaking rhythm is better. The stool-pulling motion and Donald Duck's leg flailing are both clean and natural.

But the finger choreography problem persists: Donald Duck's hands are shaking so much I can't even count how many fingers there are.

Hands shaking too fast — maybe that's just Kuaishou.

And again, completely ignored my "people sending gifts in a Kuaishou livestream room" instruction.

Tried Dreamina again:

Dreamina once again successfully wrapped the video in a Kuaishou livestream shell, but once again it's gibberish everywhere, Oracle characters all over the ground.

Motion fluidity is about the same as Keling AI, advantage being no finger choreography issues.

But Dreamina's most godlike move: exercising its own agency, it added some in-character dialogue for Tiger Bro, shouting something about "Slaughtermat Legion Commander," followed by some incomprehensible Old God whispers.

I mean, it's kinda interesting, but as a video foundation model can you just follow the goddamn prompt? Generate the video I asked for! #GaziSmash.jpg

Final test — this time I deliberately picked something extremely difficult to execute, complex in both scene and action.

I simplified it somewhat and sent the prompt to Keling AI.

Not the full prompt

Final product:

Pleasant surprise: this time it finally looks somewhat like a Kuaishou livestream.

But the generated audience comments flicker in and out like The Matrix digital rain — very postmodern.

And putting aside how the firecracker went off, after the explosion there's zero trace left — no debris on the ground, Tiger Bro's face isn't blackened, way too eco-friendly. Who invented progressive firecrackers?

Did a Dreamina version too:

For some reason, Dreamina's Tiger Bro and firecrackers both look P'd onto the background.

And the guy lighting the firecracker — where'd that cigarette in his hand come from? At first glance the motion flows smoothly, look closer and it's all magic tricks.

But at least there's debris after this firecracker explodes, slight edge there.

Later I wondered: has Keling AI completely cut ties with this low-end vibe, which is why all the above videos were disappointing?

So I specifically picked a high-end scene, high-end event: in a magnificently decorated auditorium, foreign entrepreneur Elon Musk holding up his phone, leaning in next to Tiger Bro for a photo.

Final product:

Not sure if anyone noticed, but in this video, besides Tiger Bro, every single person regardless of gender has Musk's face.

And the woman-Musk on the far left — that hand she's extending doesn't look human at all, looks kinda like a lizard person's hand, absolutely terrifying.

For fairness, I also tried Dreamina, but they detected Musk and straight-up refused to generate 😭

What's that supposed to mean, Musk is human but Tiger Bro isn't?

I then had Keling AI and Dreamina each make a Kweichow Moutai-style ad for "Northeast Good Juice," requesting Northeastern floral-print nail art on the hands in the ad.

Final product:

Judge for yourselves.

After these five rounds, I'm feeling like Keling AI's much-touted native 4K feature might actually be a pit they dug for themselves.

Keling AI talks a good game:

Everyone else generates 1080P first then upscales to 4K, and the algorithmic optimization process introduces errors and variations;

We're the world's first native 4K generation, every frame was 4K in the womb, so it looks sharper and more faithful.

But all of this assumes your video model's generation capability is solid to begin with.

If your baseline was already hyperrealistic 1:1 fine art, then switching to a finer brush would indeed make every pore feel like it's breathing.

But if what you were generating before was oil-painting skin texture, four fingers here and six fingers there, and lips that don't sync with the audio — then clearer footage just makes the flaws more glaring.

4K is basically a magnifying glass: it amplifies strengths and weaknesses alike.

Like these two images below — if not for native 4K output, I'd never have known how jarring Keling AI's generated backgrounds are.

Almost reaching world-model levels

I even suggest Keling AI should research a 480p foundation model, train directly on Kuaishou's unparalleled archive of wild content, then educate users on aesthetics — declare that in this noisy era, low-information video is good video, blurry AI is great AI.

That way nobody would discover all the existing problems.

But at the end of the day, Keling AI is still that small-town tough guy — does have genuinely irreplaceable strengths.

I had Elon Musk and Tiger Bro each star in ads for an EV brand and Hermès, to test Keling AI's cinematic quality and physics simulation.

First, the EV one:

While the wheels are a blur in motion and the exhaust doesn't look like real-world physics.

But the lighting and reflection changes on the car body as Musk walks by are quite realistic.

When Musk gets in the car, the seat, steering wheel, and door movements are all natural.

Not bad.

Then Hermès:

This one's actually decent. Though the lip sync is off, the camera movement is silky, the clothing drapes physically correctly, the Birkin's fine texture is clear, and the hardware has smooth light play.

And despite Tiger Bro starring, there's some sophistication before he opens his mouth to deliver lines. Suggest Hermès' back office contact us to purchase this for product marketing.

See, these two clips show that for static object scenes, Keling AI's 4K is a plus.

But as soon as there are people or other moving objects in frame, it's kinda over.

Maybe that's Keling AI's path forward: HD generation, 4K generation, but just focus on generating things that can't move.

After all, they got burned doing wild stunts before, so from now on they only do dead ones.

(Cover image generated by Keling AI, article purely human-written)

Finally — who still hasn't signed up for our internet café hackathon!

Current prize pool includes a Switch, two Sony WH-1000XM6s, ¥10,000 worth of Moonshot AI membership, and endless Kangshifu instant noodles on-site. Missing it is bleeding money 🩸