Capabilities
Text-to-Image
Also known as: AI image generation
AI that creates a picture from a written description; type what you want, and it generates a matching image.
Text-to-image AI turns your words into pictures. Describe a scene, "a red bicycle in front of an old Belgian café, morning light", and it produces an original image to match. It is one of the most eye-catching uses of generative AI, used for art, design, marketing and fun. The images are newly created, not found or copied.
Text-to-image models generate original images from a text prompt. Most are built on diffusion models: the system starts from random noise and refines it step by step into a picture that matches your description.
How words steer the image
The model has learned links between language and visuals (via shared embeddings), so it can push the image toward your words. Writing good prompts, subject, style, lighting, composition, is its own skill, a visual form of prompt engineering.
Things to keep in mind
- Fine details (hands, text in the image, exact counts) can still go wrong.
- There are real debates about training data, copyright and misuse (deepfakes).
Text-to-image systems condition a generative model, predominantly latent diffusion, on a text embedding to synthesise images, using classifier-free guidance to trade off fidelity to the prompt against diversity (e.g. DALL·E 2; Ramesh et al. 2022). Text-image alignment is learned from large caption datasets; a text encoder maps the prompt into a space the image generator attends to.
Capabilities and limits: strong on style, composition and photorealism, weaker (though improving) on precise text rendering, counting, spatial relations and compositional binding. The multimodal grounding that enables control also inherits dataset biases, and the field raises unresolved questions around training-data provenance and copyright, consent, and misuse for non-consensual or deceptive imagery (deepfakes), motivating provenance/watermarking standards. Related tasks (image editing, inpainting, text-to-video) extend the same conditioning machinery.