Definition
Diffusion-based models are a class of generative AI that create new data samples by iteratively removing noise from a random distribution. The process begins with a forward phase that slowly adds Gaussian noise to data until it becomes pure randomness, followed by a reverse phase where a neural network learns to predict and remove this noise step-by-step. This method has become highly effective for high-fidelity image, audio, and video generation, surpassing many previous generative adversarial networks in quality and stability.
Summary
A generative modeling approach that creates data by reversing a gradual noise-addition process through learned denoising steps.
Key Concepts
- forward process
- reverse process
- denoising
- latent space
Use Cases
- High-resolution image synthesis
- Text-to-image generation
- Data augmentation for medical imaging