Case study: DALL·E
Définition
DALL·E (and DALL·E 2) are modèles texte-vers-image d'OpenAI. They generate images from text prompts using diffusion models and language–image alignment.
Ils sont a leading example of multimodal generation: text in, image out. The same diffusion and conditioning ideas appear in Stable Diffusion and other open models. Use case: creative and product imagery from natural language; safety and content policies apply.
Comment ça fonctionne
Text is encoded with a language or multimodal encoder (par ex. CLIP text encoder, T5) into a text embedding. A diffusion model (par ex. UNet) is conditioned on this embedding: the denoising process is guided so the generated image matches the text. Training uses large datasets of captioned images; the model learns to associate text and image content. Sampling: start from noise, run the reverse diffusion process with the text embedding as condition, and decode to an image. Safety filters (par ex. classifier, policy) limit harmful or restricted outputs before delivery. Variants (inpainting, editing) condition on both text and an existing image or mask.
Cas d'utilisation
Text-to-image models like DALL·E are used wherever you need images generated or edited from natural language (creative, product, UI).
- Creative and marketing asset generation from text prompts
- Concept art, illustration, and conception exploration
- Product and UI mockups from natural language descriptions