{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/learning-warped-guidance-for-blind-face","title":"Learning Warped Guidance for Blind Face Restoration","arxiv_id":"1804.04829","date":"2018-04-13","proceeding":"ECCV 2018 9","authors":["Xiaoming Li","Ming Liu","Yuting Ye","WangMeng Zuo","Liang Lin","Ruigang Yang"],"abstract":"This paper studies the problem of blind face restoration from an\nunconstrained blurry, noisy, low-resolution, or compressed image (i.e.,\ndegraded observation). For better recovery of fine facial details, we modify\nthe problem setting by taking both the degraded observation and a high-quality\nguided image of the same identity as input to our guided face restoration\nnetwork (GFRNet). However, the degraded observation and guided image generally\nare different in pose, illumination and expression, thereby making plain CNNs\n(e.g., U-Net) fail to recover fine and identity-aware facial details. To tackle\nthis issue, our GFRNet model includes both a warping subnetwork (WarpNet) and a\nreconstruction subnetwork (RecNet). The WarpNet is introduced to predict flow\nfield for warping the guided image to correct pose and expression (i.e., warped\nguidance), while the RecNet takes the degraded observation and warped guidance\nas input to produce the restoration result. Due to that the ground-truth flow\nfield is unavailable, landmark loss together with total variation\nregularization are incorporated to guide the learning of WarpNet. Furthermore,\nto make the model applicable to blind restoration, our GFRNet is trained on the\nsynthetic data with versatile settings on blur kernel, noise level,\ndownsampling scale factor, and JPEG quality factor. Experiments show that our\nGFRNet not only performs favorably against the state-of-the-art image and face\nrestoration methods, but also generates visually photo-realistic results on\nreal degraded facial images.","url_abs":"http://arxiv.org/abs/1804.04829v2","url_pdf":"http://arxiv.org/pdf/1804.04829v2.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"learning-warped-guidance-for-blind-face","repo_url":"https://github.com/csxmli2016/GFRNet","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok"}}],"tasks":[{"task_slug":"blind-face-restoration","task_name":"Blind Face Restoration"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/image-super-resolution-on-vggface2-8x","task":"Image Super-Resolution","dataset":"VggFace2 - 8x upscaling","model":"GFRNet","rank_in_archive_order":2,"of":7,"metrics":{"PSNR":"24.10"},"uses_additional_data":false},{"leaderboard":"/sota/image-super-resolution-on-webface-8x","task":"Image Super-Resolution","dataset":"WebFace - 8x upscaling","model":"GFRNet","rank_in_archive_order":1,"of":7,"metrics":{"PSNR":"27.21"},"uses_additional_data":false}],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1804.04829","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}