Esta nota todavía no está traducida, así que se muestra la fuente en inglés.
Diffusion models, intuitively
Diffusion models generate by learning how to reverse noise. During training, clean data is gradually corrupted; the model learns to predict and remove the noise at each step.
The training idea
- Start with a real image, audio clip, or latent representation.
- Add a small amount of noise.
- Repeat until the sample is nearly pure noise.
- Train a neural network to predict the noise or the clean sample at a given timestep.
- At generation time, start from noise and denoise step by step.
The model does not memorize a lookup table of images. It learns a denoising direction that moves noisy samples toward the data distribution.
Key components
| Component | Role |
|---|---|
| Noise schedule | controls how quickly signal is destroyed |
| Denoising network | predicts noise, clean sample, or velocity |
| Timestep embedding | tells the model how noisy the sample is |
| Sampler | chooses the reverse denoising trajectory |
| Conditioning | injects text, image, class, audio, or layout guidance |
Why diffusion became dominant
- Stable training compared with older adversarial approaches.
- High sample quality and diversity.
- Flexible conditioning for text, image, masks, depth, pose, and more.
- Editing workflows such as inpainting and image-to-image.
Pitfall
Diffusion quality depends heavily on sampling settings. More denoising steps can help but also increase latency and cost, so product systems need quality-speed tradeoffs.
Connects to: training dynamics · latent diffusion · inference optimization