Image generation computationally synthesizes new images from learned representations or specified conditions. Diffusion models begin with random noise and iteratively denoise it into an image resembling the training distribution. section 7.4 section 10.1
A diffusion generation pipeline typically proceeds as follows:
- Set an inference schedule, such as 50 descending time steps. section 7.4
Initial noise determines the result: different tensors produce different images, while interpolating between tensors creates a visual transition between two generated images. Unconditional models generate without labels; conditional models can align outputs with labels or text prompts. section 7.4.1 section 6.4.1
Prompt-controlled generation
Stable Diffusion converts prompts such as “an astronaut in a spacesuit riding a unicorn” into coherent images. Prompt batches create image grids, while width, height, and guidance scale adjust the generation; larger batches and resolutions require more GPU memory. section 10.1
