A continuaciónSiguiente guía
Atención cruzada
Idioma IA
GUÍA visual de IA
Prompt-to-Prompt is a text-driven editing method for diffusion-generated images.
Escuche esta guíaNarrado por IA · ~2 min
It controls cross-attention maps that link prompt tokens to spatial features so edits can change selected concepts while attempting to retain other image structure; preservation is not exact or guaranteed.
Hertz et al. describe Prompt-to-Prompt as text-only editing of diffusion-generated images. The method intervenes on cross-attention maps, which connect text tokens to spatial image features during denoising. Replacing a word can modify its associated concept, adding a word can introduce a new attribute, and reweighting changes a token’s influence. The paper reports results on selected prompts and images; other parts of an image may change, so scene preservation is not perfect. The official implementation includes LocalBlend, which uses attention maps associated with selected words to construct a spatial mask. Prompt-to-Prompt edits a generated image’s diffusion trajectory. Editing a real photograph generally requires an additional inversion procedure. InstructPix2Pix is a separate model trained on generated image-edit instruction pairs to follow written edits from an input image.
Illustrative: replace one object in a generated scene and inspect whether unrelated details remain stable.
Illustrative: add a descriptive attribute to a prompt and compare the resulting composition.
Paper method: reweight attention for a selected token to adjust its influence on a generated image.
Official implementation: LocalBlend uses selected-word attention maps to form a spatial mask.
Evaluate edits with both target-change and preservation measures on the intended model, prompt types, and image set. Compare attention control with instruction-following editors under matched conditions. Any claim of perfect preservation, broad adoption, or future integration needs independent evidence.
Boletín gratuito
Tres noticias verificadas sobre IA cada mañana de lunes a viernes, explicadas con claridad. Gratis para siempre y sin anuncios.
Un correo cada día laborable. Date de baja con un clic. Nunca vendemos ni compartimos tu dirección.
Ponte a prueba
Corrección inmediata en cada respuesta y un certificado para compartir, con un ID verificable, cuando apruebas un curso.
Apoya la educación gratuita sobre IA. AI Understanding es una organización sin fines de lucro 501(c)(3): sin anuncios ni muros de pago, nunca. Hacer una donación
Prompt-to-Prompt is a text-driven editing method for diffusion-generated images. It controls cross-attention maps that link prompt tokens to spatial features so edits can change selected concepts while attempting to retain other image structure; preservation is not exact or guaranteed.
Los mapas de atención cruzada codifican dónde influye cada palabra en la imagen, por lo que reutilizarlos mantiene el diseño consistente durante las ediciones.
Manipula la atención en el momento de la inferencia y no necesita entrenamiento adicional.
Escalar los valores de atención de un token intensifica o silencia la fuerza con la que aparece ese concepto.
InstructPix2Pix is a separate conditional diffusion model trained to edit an input image from a written instruction. Its paper describes generated training examples; this does not establish that it derives from Prompt-to-Prompt.
The official implementation’s LocalBlend uses attention maps associated with selected words to construct a mask that restricts latent changes to relevant regions.
sigue aprendiendo
Más guías seleccionadas para este tema.
A continuaciónSiguiente guía
Atención cruzada
Idioma IA