Image generation creates new visual content by learning the underlying distribution of image data and sampling new, plausible images from those learned patterns. Unlike text generation, it must model spatial dependencies among pixels and their arrangement across the whole image. section 10.0
A typical generative approach develops through these ideas:
- VAEs and GANs established early successful architectures for generating images. section 10.0
- Diffusion models gradually add scheduled noise to an image, then learn to reverse that process. section 10.10
- A neural network predicts the noise added at each timestep, allowing sampling to subtract noise step by step. section 10.10
From pixels to plausible images
A diffusion model can start with a highly corrupted image or noise and repeatedly estimate what should be removed. Sinusoidal timestep embeddings help represent where the sample is in this process. This principle underlies systems such as DALL-E and Stable Diffusion. section 10.10
