Image generation model

Image generation models turn a text description, and often one or more input images, into a new image. Most also support editing: changing part of an image, extending it, or restyling it while keeping the rest intact. They're offered by providers such as OpenAI and Google, by specialist image companies, and as open-weight models you can host. Anthropic doesn't offer image generation, so a Claude-based product calls a separate provider for this step.

In products, the hard part is consistency, not one impressive image. Users expect the tenth image to match the first, brand colors to hold, and edits not to disturb what they liked. Prompt templates with fixed style blocks, reference images, and a text model that turns user requests into structured prompts all help. Generation is slow compared with text and priced per image or by resolution, so run it asynchronously, store results in object storage, and reuse outputs where you can.

Image features attract abuse, so moderate both the prompt and the result, and decide in advance what you'll refuse, such as real people or explicit content. Check the provider's terms for commercial use, preserve the provenance metadata many providers attach, and label generated images visibly.