what is a diffusion model
A diffusion model is a type of generative AI model that learns to create new content, such as images, by starting from pure random noise and gradually removing that noise in many small steps until a coherent, realistic output emerges — the reverse of a training process that gradually adds noise to real examples.
```flowchart
{
"rows": [
{ "type": "nodes", "items": [
{ "title": "Training: noise gradually added to real images", "sub": "model learns to predict and reverse each noise step", "tone": "blue" }
]},
{ "type": "arrow" },
{ "type": "group", "title": "Generation starts from pure random noise", "items": [
{ "title": "Model removes a little noise at each step", "sub": "repeated many times, guided by what it learned", "tone": "orange" }
]},
{ "type": "arrow" },
{ "type": "nodes", "items": [
{ "title": "Coherent, realistic output emerges", "sub": "built up gradually rather than generated all at once", "tone": "green" }
]}
]
}
```
**A diffusion model learns to generate content by first mastering the reverse of a much simpler process: gradually adding random noise to real examples until they become pure noise.** During training, the model is repeatedly shown real images with increasing amounts of random noise added at each step, and learns to predict how to remove a bit of that noise at each stage; once trained, generating something new simply runs this learned denoising process in reverse, starting from pure random noise and gradually refining it, step by step, into a coherent final output.
```svg
```
```svg
```
| Aspect | Direct generation | Diffusion model |
|---|---|---|
| Generation process | Output produced in one pass | Built gradually over many denoising steps |
| Training signal | Direct output examples | Learning to reverse a noise-adding process |
| Output quality/control | Varies by approach | Often high quality, controllable via guidance |
| Common use | Various generative approaches | Image generation, increasingly other media |
**Diffusion models have become especially prominent in image generation, though the same underlying approach has expanded to other types of content as well.** While diffusion models first gained widespread attention for producing high-quality, realistic generated images, the same core noise-and-denoise approach has since been extended to generating other content types, including audio and video, demonstrating that the underlying technique isn't inherently limited to images alone.
**Guidance techniques let a diffusion model's generation process be steered toward a desired outcome, such as matching a text description, rather than producing a purely random result.** By incorporating additional guidance signals, most commonly a text description of the desired output, during the step-by-step denoising process, diffusion models can be steered to generate content matching a specific prompt rather than simply producing an arbitrary realistic-looking result — this guided generation capability is central to how most popular text-to-image diffusion systems work.
**Running a diffusion model's many sequential denoising steps is computationally intensive, which has driven significant research into reducing the number of steps needed without sacrificing quality.** Because generating one output can require dozens or more sequential denoising steps, each requiring a model computation, diffusion model inference can be notably slower than simpler generation approaches — this cost has motivated substantial research into techniques that achieve comparable quality with meaningfully fewer denoising steps.
Read the diffusion model through a sculpting-from-noise lens: rather than producing a finished output in one shot, it starts from pure randomness and, guided by everything it learned reversing that same process during training, chips away at that randomness step by step until a coherent, realistic result gradually takes shape.