What is a diffusion model?
AITube Glossary · Updated
A diffusion model is a generative AI model that creates images or video by starting from random noise and removing it step by step until a picture matching the prompt emerges. It learns this by being trained to reverse a process that gradually adds noise to real images and videos.
How does a diffusion model work?
Training has two halves. In the forward process, real images or video frames are corrupted with a little more noise at each step until nothing recognizable is left. The model learns the reverse: given a noisy input and a description, predict the noise so it can be removed.
Generation runs only the reverse half. It starts from pure noise and, over a series of steps, removes noise while the text prompt guides what appears. Most modern systems work in a compressed latent space rather than on pixels, which makes them much faster.
Steps, samplers and seeds
More denoising steps generally add detail but take longer. A sampler, such as DDIM or Euler a, is the method used to move from one step to the next; some reach good results in fewer steps. The seed fixes the starting noise, so the same seed, prompt and settings reproduce the same output.
Diffusion models for video
Video models extend the same idea across time: they denoise all the frames of a clip together, sharing information between frames so motion and identity stay stable. Many current video models use a transformer as the network that does the denoising, an approach often called a diffusion transformer.
Why it matters to creators
Every frame of every step costs computing time, which is why video generation is priced by model, length and resolution. On AITube, larger models and longer clips cost more credits for the same reason: they do more denoising work.
Frequently asked questions
Are all AI video generators diffusion models?
Most leading ones are, frequently with a transformer backbone. Other approaches exist, including autoregressive models that predict video piece by piece, but diffusion is the dominant method.
How is a diffusion model different from a GAN?
A GAN produces an image in one pass and is trained by pitting a generator against a discriminator. Diffusion models build the output over many small steps, which is slower but trains more reliably and produces more varied, higher-quality results.
Why is AI video slower to generate than an image?
A clip contains dozens or hundreds of frames that must be denoised together and kept consistent with each other, so it takes far more computation than a single image.
Related terms
Try it on AITube
Generate AI video in AI Studio, or turn one idea into a finished vertical Short, with voiceover and captions, in StoryShort.