{"id":{"repo_id":"uiuc","oai_identifier":"oai:www.ideals.illinois.edu:2142/102527"},"canonical_url":"https://search.dev.ndltd.org/etd/uiuc/oai:www.ideals.illinois.edu:2142/102527","repository":{"repo_id":"uiuc","name":"University of Illinois - Urbana-Champaign","base_url":"https://www.ideals.illinois.edu/oai-pmh"},"display":{"title":"Improvement and measurement of neural style transfer","abstract":"Style transfer methods produce a transferred image which is a rendering of a content image in the manner of a style image. There is a rich literature of variant methods. We seek to understand how to improve style transfer: in particular, there is some evidence that cross-layer losses are helpful, and some evidence that optimization problems might present difficulties. To do so requires quantitative evaluation procedures, but current evaluation is qualitative, mostly involving user studies. We describe a novel quantitative evaluation procedure. Our procedure relies on two statistics: the Effectiveness (E) statistic measures the extent that a given style has been transferred to the target, and the Coherence (C) statistic measures the extent to which the original image's content is preserved. Our statistics are calibrated to human preference: targets with larger values of E (resp C) will reliably be preferred by human subjects in comparisons of style (resp. content). We use these statistics to investigate relative performance of a number of recent style transfer methods, revealing a number of intriguing properties. {Our experiments pool multiple style transfers from many different styles to many different content images using many different style weights, allowing us to make general statements about what influences style transfer. }Admissible methods lie on a Pareto frontier (i.e. improving E reduces C, or vice versa). Three methods are admissible: Universal style transfer produces very good C but weak E; modifying the optimization used for Gatys' loss produces a method with strong E and strong C; and a modified cross-layer method has slightly better E at strong cost in C. While the histogram loss improves the E statistics of Gatys' method, it does not make the method admissible. Surprisingly, style weights have relatively little effect, and most variability in transfer is explained by the style itself (meaning experimenters can be misguided by selecting styles).","abstract_html":"Style transfer methods produce a transferred image which is a rendering of a content image in the manner of a style image. There is a rich literature of variant methods. We seek to understand how to improve style transfer: in particular, there is some evidence that cross-layer losses are helpful, and some evidence that optimization problems might present difficulties. To do so requires quantitative evaluation procedures, but current evaluation is qualitative, mostly involving user studies. We describe a novel quantitative evaluation procedure. Our procedure relies on two statistics: the Effectiveness (E) statistic measures the extent that a given style has been transferred to the target, and the Coherence (C) statistic measures the extent to which the original image's content is preserved. Our statistics are calibrated to human preference: targets with larger values of E (resp C) will reliably be preferred by human subjects in comparisons of style (resp. content). We use these statistics to investigate relative performance of a number of recent style transfer methods, revealing a number of intriguing properties. {Our experiments pool multiple style transfers from many different styles to many different content images using many different style weights, allowing us to make general statements about what influences style transfer. }Admissible methods lie on a Pareto frontier (i.e. improving E reduces C, or vice versa). Three methods are admissible: Universal style transfer produces very good C but weak E; modifying the optimization used for Gatys' loss produces a method with strong E and strong C; and a modified cross-layer method has slightly better E at strong cost in C. While the histogram loss improves the E statistics of Gatys' method, it does not make the method admissible. Surprisingly, style weights have relatively little effect, and most variability in transfer is explained by the style itself (meaning experimenters can be misguided by selecting styles).","abstract_has_math":false,"creators":["Yeh, Mao-Chuang"],"institution":"University of Illinois at Urbana-Champaign","degree_name":"M.S.","degree_level":"Thesis","degree_discipline":"Computer Science","degree_department":null,"school":null,"contributors":["Forsyth, David A."],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2019,"date_issued":"2019-02-06T19:36:45Z","date_published":"2019-02-06T19:36:45Z","updated_at":"2026-07-22T22:24:42Z","subjects":["style transfer","gram matrix","texture synthesis"],"languages":["en"],"rights":["Copyright 2018 Mao-Chuang Yeh"],"rights_urls":[],"identifier_entries":[]},"links":{"outbound_url":"http://hdl.handle.net/2142/102527","outbound_label":"Handle","outbound_source":"dc:identifier"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor","label":"Contributor","values":["Forsyth, David A."]},{"key":"dc:creator","label":"Author","values":["Yeh, Mao-Chuang"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date","label":"Dc Date","values":["2019-02-06T19:36:45Z","2018-12-14","2018-12"]},{"key":"dc:type","label":"Dc Type","values":["text"]},{"key":"thesis:degree_discipline","label":"Discipline","values":["Computer Science"]},{"key":"thesis:degree_level","label":"Degree Level","values":["Thesis"]},{"key":"thesis:degree_name","label":"Degree Name","values":["M.S."]},{"key":"thesis:institution_name","label":"Thesis Institution Name","values":["University of Illinois at Urbana-Champaign"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["style transfer","gram matrix","texture synthesis"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language","label":"Dc Language","values":["en"]},{"key":"dc:rights","label":"Dc Rights","values":["Copyright 2018 Mao-Chuang Yeh"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["http://hdl.handle.net/2142/102527"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description","label":"Description","values":["Style transfer methods produce a transferred image which is a rendering of a content image in the manner of a style image. There is a rich literature of variant methods. We seek to understand how to improve style transfer: in particular, there is some evidence that cross-layer losses are helpful, and some evidence that optimization problems might present difficulties. To do so requires quantitative evaluation procedures, but current evaluation is qualitative, mostly involving user studies. We describe a novel quantitative evaluation procedure. Our procedure relies on two statistics: the Effectiveness (E) statistic measures the extent that a given style has been transferred to the target, and the Coherence (C) statistic measures the extent to which the original image's content is preserved. Our statistics are calibrated to human preference: targets with larger values of E (resp C) will reliably be preferred by human subjects in comparisons of style (resp. content). We use these statistics to investigate relative performance of a number of recent style transfer methods, revealing a number of intriguing properties. {Our experiments pool multiple style transfers from many different styles to many different content images using many different style weights, allowing us to make general statements about what influences style transfer. }Admissible methods lie on a Pareto frontier (i.e. improving E reduces C, or vice versa). Three methods are admissible: Universal style transfer produces very good C but weak E; modifying the optimization used for Gatys' loss produces a method with strong E and strong C; and a modified cross-layer method has slightly better E at strong cost in C. While the histogram loss improves the E statistics of Gatys' method, it does not make the method admissible. Surprisingly, style weights have relatively little effect, and most variability in transfer is explained by the style itself (meaning experimenters can be misguided by selecting styles).","Submission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2019-02-05 without embargo terms","The student, Mao-Chuang Yeh, accepted the attached license on 2018-12-14 at 16:39.","The student, Mao-Chuang Yeh, submitted this Thesis for approval on 2018-12-14 at 16:48.","This Thesis was approved for publication on 2018-12-14 at 16:56.","DSpace SAF Submission Ingestion Package generated from Vireo submission #13334 on 2019-02-05 at 11:16:19","Made available in DSpace on 2019-02-06T19:36:45Z (GMT). No. of bitstreams: 2 YEH-THESIS-2018.pdf: 23875746 bytes, checksum: e2106674506799317cd7a062f48ea7ff (MD5) LICENSE.txt: 4211 bytes, checksum: d59b85bfd93616ca18a0b99198ff772b (MD5) Previous issue date: 2018-12-14"]},{"key":"dc:format","label":"Dc Format","values":["application/pdf"]},{"key":"dc:title","label":"Title","values":["Improvement and measurement of neural style transfer"]}]}],"canonical_facts":{"dc:contributor":["Forsyth, David A."],"dc:creator":["Yeh, Mao-Chuang"],"dc:date":["2019-02-06T19:36:45Z","2018-12-14","2018-12"],"dc:description":["Style transfer methods produce a transferred image which is a rendering of a content image in the manner of a style image. There is a rich literature of variant methods. We seek to understand how to improve style transfer: in particular, there is some evidence that cross-layer losses are helpful, and some evidence that optimization problems might present difficulties. To do so requires quantitative evaluation procedures, but current evaluation is qualitative, mostly involving user studies. We describe a novel quantitative evaluation procedure. Our procedure relies on two statistics: the Effectiveness (E) statistic measures the extent that a given style has been transferred to the target, and the Coherence (C) statistic measures the extent to which the original image's content is preserved. Our statistics are calibrated to human preference: targets with larger values of E (resp C) will reliably be preferred by human subjects in comparisons of style (resp. content). We use these statistics to investigate relative performance of a number of recent style transfer methods, revealing a number of intriguing properties. {Our experiments pool multiple style transfers from many different styles to many different content images using many different style weights, allowing us to make general statements about what influences style transfer. }Admissible methods lie on a Pareto frontier (i.e. improving E reduces C, or vice versa). Three methods are admissible: Universal style transfer produces very good C but weak E; modifying the optimization used for Gatys' loss produces a method with strong E and strong C; and a modified cross-layer method has slightly better E at strong cost in C. While the histogram loss improves the E statistics of Gatys' method, it does not make the method admissible. Surprisingly, style weights have relatively little effect, and most variability in transfer is explained by the style itself (meaning experimenters can be misguided by selecting styles).","Submission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2019-02-05 without embargo terms","The student, Mao-Chuang Yeh, accepted the attached license on 2018-12-14 at 16:39.","The student, Mao-Chuang Yeh, submitted this Thesis for approval on 2018-12-14 at 16:48.","This Thesis was approved for publication on 2018-12-14 at 16:56.","DSpace SAF Submission Ingestion Package generated from Vireo submission #13334 on 2019-02-05 at 11:16:19","Made available in DSpace on 2019-02-06T19:36:45Z (GMT). No. of bitstreams: 2 YEH-THESIS-2018.pdf: 23875746 bytes, checksum: e2106674506799317cd7a062f48ea7ff (MD5) LICENSE.txt: 4211 bytes, checksum: d59b85bfd93616ca18a0b99198ff772b (MD5) Previous issue date: 2018-12-14"],"dc:format":["application/pdf"],"dc:identifier":["http://hdl.handle.net/2142/102527"],"dc:language":["en"],"dc:rights":["Copyright 2018 Mao-Chuang Yeh"],"dc:subject":["style transfer","gram matrix","texture synthesis"],"dc:title":["Improvement and measurement of neural style transfer"],"dc:type":["text"],"thesis:degree_discipline":["Computer Science"],"thesis:degree_level":["Thesis"],"thesis:degree_name":["M.S."],"thesis:institution_name":["University of Illinois at Urbana-Champaign"]},"updated_at":"2026-07-22T22:24:42Z"}