Beyond tasks tied to textual information (writing articles, writing code, doing analysis), AI tools are also increasingly used to process other kinds of information — also called "modality" : generating photorealistic photos from text, imitating human speech or composing music, and even producing cinematic-grade video from a single sentence. This is collectively known as "multi-modal" AI systems.

These capabilities and tools look very different:

·ChatGPT and DeepSeek generate text

·Midjourney and Doubao generate images

·Azure AI Speech and ElevenLabs generate speech

·Seedance, Runway and Kling generate video

But the core behind them is in fact converging.

As the first article in this series, we take image generation as our entry point to break down how modern multimodal AI systems work.

We first focus on the most popular and most curiosity-provoking question:

AI — how does it actually draw images?

1. AI drawing: is it really “drawing”?

When many people first encounter AI drawing, they get an intuition:

Does AI paint an image stroke by stroke, like a human?

In fact, it does not.

Modern mainstream AI drawing systems work more like this:

They gradually “carve” an image out of a mass of random noise.

This process is called:

Diffusion (diffusion model)

It is also the core technical approach of most advanced image-generation systems today. It covers most mainstream

AI drawing models, which all fall into this category.

2. Why did diffusion models become mainstream?

Before diffusion models appeared, the field of AI drawing relied mainly on GANs (generative adversarial networks).

Simply put:

A GAN is like a painter and a judge competing with each other: the painter generates images, the judge decides what is real and what is fake.

This approach once achieved huge success, but it has two problems:

  • Unstable training

  • Prone to getting stuck in a fixed style

Diffusion models, by contrast, take a completely different approach:

Image → keep adding noise → becomes pure noise

Then they learn the reverse process:

Pure noise → gradual denoising → restoring the image

This method is more stable to train and makes it easier to generate high-quality, diverse images.

So it gradually became the industry standard.

3. AI — why not process images directly?

This brings in the most important concept in modern image generation:

Latent space

A high-resolution image carries an enormous amount of information.

For example:

1024 × 1024 pixels

It contains more than 3 million colour values.

If AI processed these pixels directly, the computational cost would be extremely high.

So researchers came up with an idea:

Compress first.

Something like:

Original image → compression → visual summary → processing → decompression → final image

The space where this “visual summary” lives is called: Latent Space (latent space)

4. What is latent space?

You can think of it as:

To AI, an image is no longer a pile of pixels but a set of abstract visual features.

For example, when a person sees a cat:

We do not remember every pixel.

Instead we remember:

  • Ears
  • Eyes
  • Fur
  • Pose
  • Colour

AI does something similar.

It compresses an image into:

  • Shape information

  • Texture information

  • Lighting information

  • Colour information

  • Spatial structure

and other abstract representations.

So modern diffusion models do not actually work in pixel space but in latent space.

That is also why AI drawing today can maintain high efficiency while guaranteeing quality.

5. AI — how does it understand objects in an image?

This brings in yet another important concept:

Patch (image patch)

Modern AI usually does not view a whole image at once.

Instead, it cuts the image into many small patches.

For example:

One large image → many small squares

Each small square becomes a numerical vector.

Just like:

Text → words

Images → image patches

To AI:

Image patches are like the “words” of the image world.

6. If it is cut into small patches, how does AI know it is a face?

This is a question many people have.

For example:

  • Eyes

  • Nose

  • Mouth

may be scattered across different image patches.

How does AI know they belong to the same face?

The answer is:

Attention (attention mechanism)

This is also the core technology behind LLMs.

The attention mechanism lets a top-left image patch and a bottom-right image patch exchange information with each other.

So the model can discover that:

Eyes + nose + mouth = a face

and ultimately build an understanding of the whole scene.

That is also why AI today can handle complex scenes and relationships between people.

7. AI drawing and LLMs — what do they have in common?

Many people assume:

LLMs are one technology

AI drawing is another

In fact, they are increasingly similar.

Their common ground includes:

1. Both need massive amounts of data

LLMs learn from internet text.

Image models learn from vast numbers of images and captions.

2. Both use Transformer

Nearly all modern large models are built on the Transformer architecture.

3. Both use Embedding

Text becomes a vector.

Images become vectors too.

What AI ultimately processes are points in a numerical space.

4. Both need later-stage alignment

After "pre-training" is complete, through:

  • Human feedback
  • Preference learning
  • Safety rules

this "post-training" continues to improve model performance.

8. AI drawing and LLMs: the biggest difference

Although they are increasingly similar, their core generation methods differ.

LLM: continuing the text

Its way of working can be understood as:

Predicting the next word

For example:

The weather in Réunion is → good

Repeating over and over.

and a complete article takes shape.

AI drawing: carving

Diffusion models, by contrast:

Random noise → removing a bit of noise → removing a bit more noise → forming an image

So:

A large language model is “continuing a story”, while a diffusion model is “carving an image”.

This is the single most important sentence for understanding the difference between the two.

9. How does an AI image come into being?

When a user enters:

A golden retriever running under a sunset

A modern system roughly goes through the following steps:

Step one: understand the request

The language model analyses the user's intent.

Identifying:

  • Golden retriever
  • Beach
  • Running
  • Sunset
  • Photographic style

Step two: generate random noise

The system first creates a completely random “snowflake image”.

Step three: gradual denoising

The model keeps judging:

Which parts do not look like the target image?

and then corrects them.

After dozens of iterations:

Sky → sea → the dog's outline → fur details

gradually emerge.

Step four: decode into a high-resolution image

The result in latent space is finally restored into a real image.

The finished image the user sees is born from this.

10. Future trend: AI does more than draw

Past image-generation systems can be understood as:

Text → image

while today's frontier systems are increasingly close to:

User request → multimodal large-model understanding → scene planning → image-model rendering → final result

In other words:

AI is evolving from a “drawing tool” into a “creative assistant”.

Future images, speech, video and even 3D content may all be generated uniformly by a single multimodal system.

That is also why understanding the principles of image generation becomes an important starting point for understanding the whole multimodal AI era.

Summary of this article

Just remember these four sentences:

  • Diffusion is the mainstream technology of modern AI drawing.

  • AI does not process pixels directly but works in latent space (Latent Space).

  • Attention lets AI understand the relationships between different image regions.

  • An LLM continues text, while a diffusion model carves an image out of noise.