Image generation turns a text prompt (and sometimes a reference image) into new pixels. Modern systems are usually diffusion or related generative models, wrapped in product UIs with safety filters and style controls.
Intuition
You describe “a watercolor fox under streetlights.” The model starts from noise (or a latent) and iteratively denoise toward an image that matches the prompt embedding. Editing tools (“inpaint this region”) constrain which pixels may change.
Practical knobs (product view)
Prompt / negative prompt — what to include and avoid.
Seed — reproducibility.
Guidance / CFG — how strongly to follow the prompt vs stay natural.
Resolution / steps — quality vs latency/cost.
What goes wrong
Brand / likeness misuse and copyright risk.
Unsafe or biased imagery if filters are weak.
Prompt injection via images in multimodal apps (treat untrusted images carefully).
One-line summary
Text-to-image models synthesize pictures from prompts; treat them as creative tools that still need safety and rights review.
Key terms
Diffusion — generate by iterative denoising from noise.
Latent — compressed space where many generators operate for speed.
Guidance — strength of prompt adherence during sampling.