From Prompt to Pixel: The Story Behind Viral AI Videos

Saturday night, Truong shared a video in our group chat.

"Everyone check this out. This video is 100% AI-made. No actors, no cameramen, nothing. Just a single prompt."

The video was 30 seconds long. A girl walking on a beach at sunset. Wind blowing through her hair. Sunlight shimmering on her skin. She turns to look at the camera and says: "Do you want to come with me?"

Perfect lip-sync. Natural voice. If no one had said anything, I would have thought it was real.

Hieu replied: "Veo3 from Google, right? Just came out last week. Insane."

Minh chimed in: "But I don't understand. Why did all this explode right when ChatGPT came out? What does LLM have to do with video generation?"

Good question. And the answer is more complex than Minh thought.

The big picture: Generative AI

Before diving deep, let's step back and look at the big picture.

AI splits into two major branches:

Discriminative AI - AI that recognizes, classifies. Takes input, classifies or predicts. This is the "old" AI, been around for a while.

Real examples you encounter daily:

• Unlock your phone with your face → AI recognizes that's you
• Gmail automatically filters spam → AI classifies emails
• YouTube recommends videos → AI predicts what you'll like
• Uber/Grab calculates fare → AI predicts time and distance

Generative AI - AI that creates. Takes input, creates something new. This is what's exploding now.

And within Generative AI, there are many branches:

• Gen text → LLM (Claude, ChatGPT) - write emails, code, answer questions
• Gen images → Diffusion Model (Midjourney, DALL-E) - create posters, avatars, concept art
• Gen video → Video Diffusion (Sora, Veo3) - create ads, TikTok clips
• Gen audio → (ElevenLabs, Suno) - clone voices, create music
• Gen code → also LLM - write functions, debug, refactor

ChatGPT, Claude, Midjourney, Sora - all are Generative AI. They're siblings in the same family.

LLM is just the most famous child, specializing in text. But Generative AI existed before LLM - GAN generating faces since 2014 was already Generative AI.

Understanding this helps you see: everything exploding at once isn't coincidence. They're from the same family.

Before ChatGPT

To understand why everything exploded at once, we need to rewind a bit.

In 2014, a new technique was born: GAN (Generative Adversarial Network). The idea was simple but genius: two neural networks competing against each other. One tries to create fake images. The other tries to distinguish real from fake. They fight continuously, and both get better over time.

GAN could generate human faces. But there was a big problem: you couldn't tell it what you wanted.

Imagine hiring a deaf-mute painter. They paint beautifully, but you can't tell them "paint me a blonde girl in a red dress." You can only gesture with your hands, or show them reference images. Very hard to communicate.

GAN was the same. Didn't understand language. Had to adjust technical parameters to get different results. Only tech people could use it.

And the quality? Oh boy. You can Google "GAN face fail" to see. Distorted faces. Extra teeth or missing teeth. Eyes looking in different directions. Sometimes 6 fingers, sometimes 4. A few seconds of looking and you'd know it was fake.

That's why you didn't see "AI images" going viral before 2022. It wasn't good enough yet.

The translator appears

In 2021, OpenAI released something called CLIP.

CLIP doesn't create images. It does something else: translates between language and images.

Imagine you have a translator. This person can look at an image and describe it in words. Conversely, you say a sentence, and they can imagine the corresponding image.

CLIP is that translator. It was trained on hundreds of millions of image-text pairs from the internet. See a beach photo, it knows that's "beach." Hear the word "sunset," it knows what that scene looks like.

Think of CLIP like Google Translate, but instead of translating English to Vietnamese, it translates between "human language" and "pixel language." When you type "girl in red dress standing in rain," CLIP converts that sentence into a numeric form - a vector - that image generation models can understand.

And CLIP was built on Transformer - the same architecture as GPT, Claude, and every modern LLM.

This is the most important piece of the puzzle. It explains why advances in LLM also pull advances in image generation - they share the same "engine."

When the translator meets the artist

Around the same time, another technique was rising: Diffusion Model.

The idea behind Diffusion Model is strange: take a beautiful image, gradually add noise, until it becomes a mess of random pixels. Then train a model to go backwards - from noise to image.

Like you have a lump of clay. You shape it into a cat. Then you smash it flat, roll it into a ball, turn it into a shapeless blob. The model learns to look at that shapeless blob and sculpt it back into a cat.

Diffusion Model creates images much better than GAN. More detailed. Fewer errors. But it still had the same problem: couldn't understand language.

Until someone thought: "Why not combine CLIP with Diffusion Model?"

CLIP understands language. Diffusion Model creates beautiful images. What if we put them together?

DALL-E 2. Stable Diffusion. Midjourney. All born from this idea.

You write a prompt: "girl walking on beach at sunset."

CLIP reads the prompt, converts it into a mathematical representation that machines understand. Diffusion Model receives that representation, then gradually "denoises" from random pixels into the image you described.

For the first time, anyone could create images just by typing text.

Why prompts matter

Truong often shares videos of content creators talking about how this prompt is better than that prompt. Minh wondered: "Are prompts really that important? Or is it just marketing?"

They really are important.

Remember CLIP? The "translator" between language and images? The prompt is what you tell that translator.

If you speak vaguely, the translator understands vaguely, miscommunicates to the artist, the image comes out wrong.

If you speak in detail, clearly, in the "language" the translator is familiar with, they communicate accurately, the image comes out right.

Example:

"Beautiful video" → CLIP doesn't know what you want. Random result.

"Cinematic shot, golden hour, young woman walking barefoot on wet sand, waves in background, slow motion, shallow depth of field" → CLIP understands every detail. Result close to what you wanted.

Words like "cinematic," "golden hour," "shallow depth of field" - they appear very often in the data CLIP was trained on. So it understands their exact meaning.

Vague words like "beautiful," "nice," "cool" - they could map to millions of different images in the data. So results are random.

Prompt engineering isn't marketing. It's how to communicate effectively with AI.

From images to video

Image generation was impressive enough. But video is what's truly shocking now.

Sora. Veo3. Runway. Kling.

Videos tens of seconds long. Smooth motion. Natural lighting. Accurate physics. And now - perfect lip-sync with voice.

Where have you already seen AI video without knowing?

• Short ads on Facebook/Instagram - many are now 100% AI-made
• TikTok videos with "turn yourself into anime character" effects
• Product review clips with virtual models
• Music videos with MVs that no one actually filmed

Things that used to require an entire film crew can now be done by one person with a laptop.

Fundamentally, video generation is image generation extended with the time dimension.

Instead of denoising into 1 image, the model denoises into a sequence of many consecutive frames. Instead of just learning "what beautiful images look like," the model must also learn "what natural movement looks like."

Water must flow according to physics. Hair must blow in the wind direction. People walking must have correct center of gravity. Mouths speaking must match the audio.

The model was trained on millions of videos from the internet. It learned all those rules. No one hardcoded them. It learned from data.

And when the model is big enough, trained on enough data, with enough compute - it starts creating things indistinguishable from reality.

Answering Minh's question

Minh asked the right question. Image generation isn't LLM. Diffusion Model isn't GPT. So why did they explode at the same time?

First: shared roots.

Transformer - the architecture behind GPT, Claude - is also the architecture behind CLIP. The text encoder in Stable Diffusion, in Midjourney, in Sora, all use Transformer. They're siblings, born from the same revolution.

Second: Scaling Law.

People discovered: keep increasing data, model size, compute, and quality increases. This discovery was true for LLM, and also true for image/video generation. When OpenAI proved this with GPT-3, the whole industry poured money into scaling everything up.

Third: the money flow.

ChatGPT went viral in late 2022. Hundreds of millions of users in weeks. For the first time, AI proved it had massive commercial potential.

Investment money poured in. Billions of dollars. Not just into LLM, but into all of AI. Midjourney, Runway, Stability AI - all benefited. Money buys GPUs. GPUs train bigger models. Bigger models mean better quality.

It wasn't coincidence. It was a domino effect.

Not just images and video

And here's the interesting thing: that domino effect didn't stop at image and video generation.

Audio generation also exploded. ElevenLabs clones voices with just 30 seconds of sample. Suno creates complete songs from prompts. You can listen to "songs written and sung by AI" that are indistinguishable from real singers.

Coding AI too. GitHub Copilot, Cursor, Claude Code - all use LLM to write code. In 2022 it was "suggesting lines of code," by 2025 it could "write entire features from scratch."

And the craziest part? Robotics.

Early 2026, Unitree (China) released videos of humanoid robots doing martial arts. Not the stiff robotic movements of old. Real kung fu - spinning kicks, high jumps, perfect balance. Boston Dynamics finally has a worthy competitor.

Why is robotics also exploding at the same time?

Because robots need a "brain" to process their environment. That brain now uses Transformer - the same architecture as GPT. Robots need to learn from data - videos of people practicing martial arts, movement data. Scaling Law applies here too: more data, bigger model, smarter robot.

Everything is connected. LLM, image generation, video generation, audio generation, robotics - all riding the same technological wave.

Minh listened, was quiet for a moment, then said: "So the entire AI world is evolving together."

Exactly. And the speed of that evolution is what's truly scary.

The scary pace of development

Hieu made a good observation: "Every year it jumps a terrifying level."

Looking back:

2022: Image generation still had warped hands, wrong fingers. Video barely existed.

2023: Images were already very good. Video started appearing but was still jittery, wrong physics.

2024: Video got smoother. But lip-sync was still off. Voices still sounded fake.

2025: Veo3 released. Video with natural voice. Accurate lip-sync. Vivid facial expressions. Nearly indistinguishable from real.

To put it in perspective: from "what is that, the hands are all messed up" to "wait, is this a real person or AI?" took only 3 years.

Three years. The length of a college degree. The time to pay off a motorbike loan. And technology jumped from meme material to masterpiece.

One leap every year. And the pace seems to be accelerating, not slowing.

Things to think about

That night, after watching the video, the whole group went quiet for a moment.

Then An - the team's QC lead - spoke up:

"I just watched a video made by AI. Beautiful like a movie. But I'm wondering: if anyone can create content like this, does content still have value?"

No one could answer that question.

On one hand, the barrier to creation is nearly zero now. Before, making an ad clip required filming, actors, post-production - tens of thousands of dollars. Now one person sitting at home writing prompts can produce high-quality video.

Online seller who doesn't know design? Now can create product posters in 30 seconds. Startup without marketing budget? Now can make ad videos without hiring an agency. Individual wanting to start YouTube? Now can create professional thumbnails without learning Photoshop.

On the other hand, when everyone can do it, content floods everywhere. "Oversaturated" as Truong would say. Most of it looks the same. Lacking real creativity.

And other issues:

• Deepfakes are becoming perfect, nearly undetectable
• Copyright is still a gray area - model trained on others' work, who owns the output?
• Many jobs are being genuinely affected: illustrators, motion designers, voice actors...

Technology isn't good or bad. It's just a tool. But this tool is changing a lot of things, very fast.

Convergence

There's a trend forming: multimodal.

GPT-4o can chat and generate images. Gemini understands text, images, video, audio. Claude can analyze images.

Imagine this scenario: you chat with AI, say "create an ad video for my coffee shop." AI asks back "what style?" You snap a photo of your shop and send it. AI looks at the photo, understands the vintage aesthetic, then creates a video with soft jazz background music, with a voiceover describing the shop. All in one conversation.

That's not distant future. GPT-4o already does part of it. Gemini 2.0 is getting closer.

The boundary between LLM and image/video generation is blurring. The future might be a single model that does everything: understand language, create images, create video, create audio, interact in realtime.

And then, distinguishing "LLM" from "image generation" might no longer make sense. They'll just be different capabilities of the same AI.

We're living in an interesting time. And a scary one. Simultaneously.

Closing thoughts

Minh asked: "What does LLM have to do with video generation?"

The answer: a lot, and deeper than you'd think.

They were born from the same revolution: Transformer. They grew up from the same discovery: Scaling Law. They exploded from the same wave: money pouring into AI after ChatGPT proved commercial value.

LLM is the most famous child of the family. But image generation, video generation, audio generation - they're all siblings.

And it seems like they're gradually merging into one.

The prompt is the bridge between the idea in your head and the AI on the other side.

Understanding how it works won't help you use it much better. But at least you'll know who you're talking to.