We propose a reversible face de-identification method for low resolution\nvideo data, where landmark-based techniques cannot be reliably used. Our\nsolution is able to generate a photo realistic de-identified stream that meets\nthe data protection regulations and can be publicly released under minimal\nprivacy constraints. Notably, such stream encapsulates all the information\nrequired to later reconstruct the original scene, which is useful for\nscenarios, such as crime investigation, where the identification of the\nsubjects is of most importance. We describe a learning process that jointly\noptimizes two main components: 1) a public module, that receives the raw data\nand generates the de-identified stream, where the ID information is surrogated\nin a photo-realistic and seamless way; and 2) a private module, designed for\nlegal/security authorities, that analyses the public stream and reconstructs\nthe original scene, disclosing the actual IDs of all the subjects in the scene.\nThe proposed solution is landmarks-free and uses a conditional generative\nadversarial network to generate synthetic faces that preserve pose, lighting,\nbackground information and even facial expressions. Also, we enable full\ncontrol over the set of soft facial attributes that should be preserved between\nthe raw and de-identified data, which broads the range of applications for this\nsolution. Our experiments were conducted in three different visual surveillance\ndatasets (BIODI, MARS and P-DESTRE) and showed highly encouraging results. The\nsource code is available at https://github.com/hugomcp/uu-net.\n