Aller au contenu
Atlas de l'IA

Texte vers image

Transformer une phrase en image

Génération et édition d'imagesDébutant #18
entréeTexteImage

Le texte intégral est présenté en anglais ; le titre et le résumé sont localisés.

CE QUE DÉSIGNE CETTE CAPACITÉ

Maps a natural-language description directly to an image: text in, pixels out, with no annotation or reference image in between. Unlike image-to-image it has no input image at all, and unlike image editing it does not modify an existing picture but synthesises a new one from scratch.

Comment c'est fait

The dominant route is latent diffusion: the prompt is encoded into a conditioning vector, injected into a denoising network through cross-attention, denoised step by step in a compressed latent space, and decoded to pixels. Conditioning was later extended by classifier-free guidance, which amplifies prompt adherence by contrasting the conditioned and unconditioned directions. Training uses vast image–text pairs, and DALL·E 3 coupled prompt rewriting with generation in one pipeline, markedly improving adherence to long prompts.

Produits représentatifs

7

Organisations concernées

Usages typiques

  • Concept design and storyboards
  • Marketing assets and illustration
  • Pre-visualisation for games and film
  • Personalised avatars and wallpapers

Comment on l'évalue

FID
Distance to the real image distribution; lower is better
CLIP score
Semantic agreement between image and prompt
Human-preference Elo
Preference ranking from pairwise comparison

Limites et points difficiles

  • Counting, precise spatial relations and ordering remain unreliable
  • Letters inside the image are frequently garbled, especially in long strings
  • Hands, limbs and object-contact points break down, often needing several resamples

Concepts sous-jacents