Context Encoders: Feature Learning by Inpainting [arXiv] (April 25 2016)

• First use of CNNs in image inpainting.
• Utilizes an adversarial loss

• Completed regions blurry.

227 x 227 x 3 for random region and 128 x 128 x 3 for center region [official torch implementation] Semantic Image Inpainting with Deep Generative Models [arXiv] (July 26 2016)

• Missing content infered by searching for closest encoding of the corrupted image in the latent image manifold.

• No end to end training.
• IMHO, generating images is harder than inpainting images because with inpainting, there is always ground truth present. So converting inpainting to the harder problem of generating images might not be the way to go.

64 x 64 x 3, arbitrary mask [tensorflow implementation] Globally and Locally Consistent Image Completion [link] (2017)

• Dilated convolutions
• 2 discriminators: one local discriminator for the completed region and one global discriminator for whole image

• Long training time.
• Poisson blending needed.
• Complex training process. Completion network is trained, then the completion network is fixed and discriminators are trained, then finally both are trained.

256 x 256 x 3 [official torch implementation. no training code] [tensorflow implementation. missing gan loss] [tensorflow implementation. has UI] Image Inpainting using Multi-Scale Feature Image Translation [arXiv] (November 23 2017)

• 2 stages: coarse prediction and refinement through feature based texture swapping
• Framework can be adapted to multi-scale

256 x 256 x 3 input, arbitrary mask Soon Context-Aware Semantic Inpainting [arXiv] (December 21 2017) Light-weight pixel context encoders for image inpainting [arXiv] (January 17 2018) High Resolution Face Completion with Multiple Controllable Attributes via Fully End-to-End Progressive Generative Adversarial Networks [arXiv] (January 23 2018)

• Conditioned on facial attributes.
• Progressive growing of GANs.
• Three new losses: attribute, feature, and boundary.

• Fails to learn low level skin features.
• Long training time.

1024 x 1024 x 3, arbitrary mask Soon Deep Structured Energy-Based Image Inpainting [arXiv] (January 24 2018) Image Inpainting for Irregular Holes Using Partial Convolutions [arXiv] (April 20 2018)

• Introduces partial convolutions, which exclude information from the mask.
• No post processing.

512 x 512 x 3, arbitrary mask Free-Form Image Inpainting with Gated Convolution [arXiv] (June 10 2018)

• Utilizes gated convolutions.
• State of the art inpainting for irregular masks.
• No post processing.

512 x 512 x 3, arbitrary mask Soon