The dominant approach: encode text into embeddings (via CLIP or T5), start with random noise, and iteratively denoise while conditioning on the text embeddings through cross-attention. Each denoising step makes the image slightly less noisy and more aligned with the prompt. After 20–50 steps (or 4–10 with flow matching), a clean image emerges. The model has learned the statistical relationship between text descriptions and image features from billions of image-caption pairs.
Beyond text prompts, modern image generation supports: image-to-image (modify an existing image), ControlNet (guide composition with edge maps, depth maps, or poses), inpainting (regenerate part of an image), and style transfer (apply the aesthetic of one image to another). These controls make image generation practical for professional workflows where "generate something random" isn't enough — you need specific compositions, poses, and layouts.
Image quality improvements come from three sources: better architectures (U-Net to DiT), better training (flow matching over diffusion), and better data (higher resolution, better captions, more diverse). Current frontier models produce photorealistic images that are difficult to distinguish from photographs, though they still struggle with: hands and fingers, text rendering, spatial relationships ("A is to the left of B"), and counting ("exactly five apples"). These remaining challenges are active research areas.