A text-to-image model synthesizes an image from a textual description. Its central challenge is connecting visual pixels with linguistic concepts, so captions teach the model what subjects, actions, and styles to associate with images. section 12.5
Useful training practices include:
- Pair images with detailed captions rather than filenames or broad labels; specificity improves the model’s ability to follow detailed prompts. section 12.5
- Use synthetic images and 3D renders when precise control over lighting, angle, composition, and labeling is valuable. section 12.5
- Account for copyright, privacy, licensing, and style-mimicry concerns when collecting image data. section 12.5
Why caption detail matters
A caption such as “A photorealistic close-up of a red-tailed hawk perched on a pine branch, its feathers ruffled by the wind, with a soft-focus lush green forest background” links the subject, action, and visual style to the image. A label such as “bird” leaves those distinctions ambiguous. section 12.5
