Atlas record: https://www.aitimeline.dev/en/events/evt-0759/
Research · November 21, 2016
pix2pix: one model for image translation
On 21 November 2016 Phillip Isola, Alexei Efros and colleagues at Berkeley showed that a conditional adversarial network with one architecture and objective translates image to image across many tasks: labels to street scenes, aerial photos to maps, day to night, edges to photos.
Why it matters
Tasks that had each needed their own algorithm became one model trained on different pairs. The loss for each task no longer had to be designed by hand either: the discriminator learns it.
The generator is a U-Net type, and a PatchGAN discriminator judges separate patches of the image. In a study on Amazon Mechanical Turk, aerial photos generated from maps fooled participants on 18.9% of trials, well above the L1 baseline.
Event record
- Event date
- November 21, 2016
- Timeline date
- Event date
- Verification
- Sources gathered automatically · September 25, 2026
- Lines
- ID
- evt-0759
The day the first version of the preprint was submitted; the figures were read in it.
Related events
Earlier
- Builds on U-Net: segmentation from a few images 2015
The pix2pix generator is a U-Net-type network with skip connections, and the paper shows the results are much better with them.
- Builds on Generative adversarial networks 2014
pix2pix is a conditional adversarial network: the generator learns to fool the discriminator, as in the GAN of 2014, which the paper cites.
- Enables Image Analogies: a filter from a pair of examples 2001links to this record
pix2pix cites Image Analogies in its first paragraph among separate image-to-image methods and again among the style models its discriminator rests on.
- Enables A texture grown from a sample 1999links to this record
pix2pix says its discriminator models the image as a Markov random field, "the common assumption in models of texture", and cites this paper first.