# What is Diffusion Model?

Also called denoising diffusion model.

A diffusion model is a generative system trained to reverse a gradual noising process. During training it learns to remove noise from corrupted data; during generation it starts from pure noise and denoises step by step until a coherent sample emerges. Diffusion is the dominant approach for image, video, and audio generation.

Training defines a forward process that adds small amounts of noise to real data over many steps until nothing but noise remains. A network is then trained to predict the noise that was added at each step. Because each individual step is a small, well-posed prediction problem, the overall task is far more stable to train than earlier adversarial approaches.

Generation runs the process backward. Starting from random noise, the model repeatedly estimates and subtracts noise, refining the sample across a scheduled number of steps. Fewer steps mean faster generation and typically lower fidelity, which is why sampler choice and step count are the main quality and speed dials exposed to users.

Conditioning steers the output. A text description is encoded and supplied to the denoising network so that each step pulls the sample toward content matching the description. Guidance techniques strengthen this pull by comparing conditioned and unconditioned predictions, trading diversity for closer adherence to the prompt. Excessive guidance produces oversaturated or rigid results.

Most production systems operate in a compressed latent space rather than on raw pixels. An autoencoder maps images into a smaller representation, diffusion runs there, and a decoder restores full resolution. This cuts compute substantially and is what made high-resolution generation practical on ordinary hardware.

## Key points

- Trained to reverse a gradual noise-adding process
- Generation starts from noise and denoises over many steps
- Step count trades generation speed against output fidelity
- Text conditioning and guidance steer the result
- Most systems diffuse in a compressed latent space

## In practice

Asked for an illustration of a bicycle on a beach, the system encodes the description, initializes a latent tensor with random noise, and runs perhaps thirty denoising steps. Early steps settle broad composition such as horizon placement and object mass; later steps resolve spokes, sand texture, and edges. A decoder then converts the final latent into pixels. Cutting the step count in half produces a recognizable but noticeably rougher image.

## Related terms

- [Generative AI](/en/glossary/generative-ai)
- [Multimodal Model](/en/glossary/multimodal-model)
- [Neural Network](/en/glossary/neural-network)
- [Inference](/en/glossary/inference)
- [Parameters](/en/glossary/parameters)

[Back to the AI Glossary](/en/glossary)