Il prossimoProssima guida
Traduzione non accoppiata di CycleGAN
IA visiva
GUIDA AI visiva
Pix2pix learns from aligned image pairs.
Ascolta questa guidaNarrato dall'IA · ~2 min
Given an outline, for example, it can generate a plausible photograph whose objects follow that outline.
During training, the generator receives an input such as a street-scene label map and predicts a color image. A U-Net carries details from its encoder to its decoder through skip connections. Road edges and building boundaries can travel through those connections while deeper layers learn how objects typically look. The discriminator sees the input alongside either the real target or a generated image. Its PatchGAN decisions evaluate local regions. That encourages convincing texture, while a separate L1 loss penalizes pixel differences from the paired target. A convincing brick texture can still appear on the wrong wall, so these two signals address different errors.
04Esempio lavorato
Draw a single black outline of a shoe. Make two target images with exactly that outline, one filled red and the other blue. Imagine both pairs appear equally often in training.
At one interior pixel, simplify the red channel to 1 in the red shoe and 0 in the blue shoe. A prediction of 0.2 has average absolute error (0.8 + 0.2) / 2 = 0.5. A prediction of 1 also has average error 0.5.
Compare a generated shoe with the outline first. Then ask whether the color looks plausible and whether anything in the input identifies the original color. Record those judgments separately.
Cosa mostra
Every red-channel prediction between 0 and 1 ties on this pixel loss. The adversarial term can favor a realistic color, but the outline alone cannot tell you which shoe the artist saw.
Translate a building facade label map into a proposed facade rendering.
Generate a shoe photograph from its outline.
Compare generated street scenes with the road and building boundaries in the input.
Newsletter gratuita
Tre storie verificate sull'IA ogni mattina nei giorni feriali, scritte in un inglese semplice. Gratuito per sempre, senza pubblicità.
Un'email ogni giorno feriale. Disiscriviti con un clic. Non vendiamo né condividiamo mai il tuo indirizzo.
Mettiti alla prova
Feedback immediato su ogni risposta e un certificato condivisibile con un ID verificabile una volta superato un corso.
Sostieni l'istruzione gratuita sull'IA. AI Understanding è un'organizzazione non profit 501(c)(3) — nessun annuncio, nessun paywall, mai. Fai una donazione
The original objective needs matching input and target images. Organize each input with its corresponding target, or consider an unpaired method such as CycleGAN if those correspondences are unavailable.
Continua a imparare
Altre guide selezionate per questo argomento
Il prossimoProssima guida
Traduzione non accoppiata di CycleGAN
IA visiva