5.2.1. Generating Images and Video from Prompts
💡 First Principle: Image and video generation is prompt-to-media: you describe what you want, optionally supply reference media to steer style or content, and the model produces the asset. The same prompt-quality discipline as text generation applies — specific, well-scoped prompts produce better assets — and platform controls (size, style, quality, safety) shape the output.
In Foundry, image generation is handled by image-generation models (the GPT-image / DALL·E family) and video generation by video-generation models (the Sora family), consumed through the same project-and-deployment pattern as text models. You send a text prompt — and optionally reference media — and receive generated image or video assets. Reference media lets you steer the result toward an existing style, subject, or composition rather than generating purely from scratch. For agentic scenarios, generation composes as a tool: an agent assembling a brochure can call image generation mid-task and place the result in the next step, exactly as it would call any other tool. You select and apply the generation controls the platform exposes — dimensions, style, quality, and the safety/content settings that gate what may be produced.
⚠️ Exam Trap: Don't answer a "produce an illustration from this description" requirement with an understanding model, and don't answer "describe this uploaded photo" with a generation model. The tell is the direction of the arrow: text→media is generation; media→text is understanding. Reference media on a generation request still means generation — it steers the output, it is not the thing being analyzed.
Reflection Question: A workflow must turn a written product description into a promotional image and also, separately, write alt-text for that image once produced. Which step is generation and which is understanding, and why must they use different model types?
