{"id":{"repo_id":"glasgow","oai_identifier":"oai:theses.gla.ac.uk:86161"},"canonical_url":"https://search.ndltd.org/etd/glasgow/oai:theses.gla.ac.uk:86161","repository":{"repo_id":"glasgow","name":"University of Glasgow","base_url":"https://theses.gla.ac.uk/cgi/oai2"},"display":{"title":"Scaling resolution via guided priors: from explicit multimodal fusion to inference-time adaptation of generative foundation models","abstract":"Achieving high spatial resolution remains a fundamental challenge in computer vision, where results often fail to adhere to physical constraints. This thesis investigates the use of guided priors to mitigate these limitations, tracing a methodological evolution from explicit multimodal fusion in supervised settings to implicit inference-time priors within generative foundation models across 2D and 3D domains. First, we address depth sensor resolution limits through explicit guidance. We propose IGAF (Incremental Guided Attention Fusion), a deep learning architecture that utilizes high-resolution RGB images to iteratively upsample depth maps. By selectively attending to cross-modal features, IGAF transfers structural guidance from the RGB domain to reconstruct fine geometric details that exceed the sensors native resolution. Shifting toward generative AI, we introduce Pixelsmith, a zero-shot framework for gigapixel image generation on a single GPU. By employing a cascading approach where lower-resolution outputs serve as explicit guidance for patch-based denoising, we leverage the foundation models learned knowledge as an implicit prior. This enables the synthesis of extreme detail while overcoming the memory constraints of standard diffusion models. Finally, we extend this paradigm to 3D with SceneHI, a framework that lifts 2D generative features into 3D space. By orchestrating a rendering / inverse rendering loop as explicit guidance to adapt implicit priors, SceneHI eliminates multi-view inconsistencies. We further integrate geometry-consistent shadows, addressing both resolution and physical plausibility constraints in complex 3D environments. Ultimately, this thesis establishes that structural guidance is an essential requirement for surpassing the physical resolution limits of sensors and foundation models across 2D and 3D domains.","abstract_html":"Achieving high spatial resolution remains a fundamental challenge in computer vision, where results often fail to adhere to physical constraints. This thesis investigates the use of guided priors to mitigate these limitations, tracing a methodological evolution from explicit multimodal fusion in supervised settings to implicit inference-time priors within generative foundation models across 2D and 3D domains. First, we address depth sensor resolution limits through explicit guidance. We propose IGAF (Incremental Guided Attention Fusion), a deep learning architecture that utilizes high-resolution RGB images to iteratively upsample depth maps. By selectively attending to cross-modal features, IGAF transfers structural guidance from the RGB domain to reconstruct fine geometric details that exceed the sensors native resolution. Shifting toward generative AI, we introduce Pixelsmith, a zero-shot framework for gigapixel image generation on a single GPU. By employing a cascading approach where lower-resolution outputs serve as explicit guidance for patch-based denoising, we leverage the foundation models learned knowledge as an implicit prior. This enables the synthesis of extreme detail while overcoming the memory constraints of standard diffusion models. Finally, we extend this paradigm to 3D with SceneHI, a framework that lifts 2D generative features into 3D space. By orchestrating a rendering / inverse rendering loop as explicit guidance to adapt implicit priors, SceneHI eliminates multi-view inconsistencies. We further integrate geometry-consistent shadows, addressing both resolution and physical plausibility constraints in complex 3D environments. Ultimately, this thesis establishes that structural guidance is an essential requirement for surpassing the physical resolution limits of sensors and foundation models across 2D and 3D domains.","abstract_has_math":false,"creators":["Tragakis, Athanasios"],"institution":"University of Glasgow","degree_name":null,"degree_level":"PhD","degree_discipline":null,"degree_department":null,"school":null,"contributors":[],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2026,"date_issued":"2026","date_published":"2026","updated_at":"2026-11-19T08:37:45Z","subjects":["QA76 Computer software"],"languages":["en"],"rights":[],"rights_urls":[],"identifier_entries":[],"person_identifier_entries":[]},"links":{"outbound_url":"https://doi.org/10.5525/gla.thesis.86161","outbound_label":"DOI","outbound_source":"dc:identifier.doi","record_url":"https://doi.org/10.5525/gla.thesis.86161","record_label":"Repository record","record_kind":"doi","record_source":"dc:identifier.doi","download_url":"https://theses.gla.ac.uk/86161/4/2026TragakisPhD.pdf","download_label":"Download PDF","download_source":"dc:identifier.uri","access_status":null,"embargo_until":null,"access_note":null},"source_record":{"url":"https://theses.gla.ac.uk/cgi/oai2?verb=GetRecord&metadataPrefix=uketd_dc&identifier=oai%3Atheses.gla.ac.uk%3A86161","prefix":"uketd_dc"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:creator","label":"Author","values":["Tragakis, Athanasios"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date","label":"Dc Date","values":["2026"]},{"key":"dc:date.issued","label":"Date","values":["2026"]},{"key":"dc:publisher.institution","label":"Dc Publisher Institution","values":["University of Glasgow"]},{"key":"dc:relation.isreferencedby","label":"Dc Relation Isreferencedby","values":["https://theses.gla.ac.uk/86161/"]},{"key":"dc:type","label":"Dc Type","values":["Thesis"]},{"key":"dc:type.qualificationlevel","label":"Dc Type Qualificationlevel","values":["PhD"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["QA76 Computer software"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language","label":"Dc Language","values":["en"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier.doi","label":"DOI","values":["10.5525/gla.thesis.86161"]},{"key":"dc:identifier.uri","label":"Identifier URI","values":["https://theses.gla.ac.uk/86161/4/2026TragakisPhD.pdf"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description.abstract","label":"Abstract","values":["Achieving high spatial resolution remains a fundamental challenge in computer vision, where results often fail to adhere to physical constraints. This thesis investigates the use of guided priors to mitigate these limitations, tracing a methodological evolution from explicit multimodal fusion in supervised settings to implicit inference-time priors within generative foundation models across 2D and 3D domains. First, we address depth sensor resolution limits through explicit guidance. We propose IGAF (Incremental Guided Attention Fusion), a deep learning architecture that utilizes high-resolution RGB images to iteratively upsample depth maps. By selectively attending to cross-modal features, IGAF transfers structural guidance from the RGB domain to reconstruct fine geometric details that exceed the sensors native resolution. Shifting toward generative AI, we introduce Pixelsmith, a zero-shot framework for gigapixel image generation on a single GPU. By employing a cascading approach where lower-resolution outputs serve as explicit guidance for patch-based denoising, we leverage the foundation models learned knowledge as an implicit prior. This enables the synthesis of extreme detail while overcoming the memory constraints of standard diffusion models. Finally, we extend this paradigm to 3D with SceneHI, a framework that lifts 2D generative features into 3D space. By orchestrating a rendering / inverse rendering loop as explicit guidance to adapt implicit priors, SceneHI eliminates multi-view inconsistencies. We further integrate geometry-consistent shadows, addressing both resolution and physical plausibility constraints in complex 3D environments. Ultimately, this thesis establishes that structural guidance is an essential requirement for surpassing the physical resolution limits of sensors and foundation models across 2D and 3D domains."]},{"key":"dc:format","label":"Dc Format","values":["pdf"]},{"key":"dcterms:accessrights","label":"Dcterms Accessrights","values":["public"]},{"key":"dc:title","label":"Title","values":["Scaling resolution via guided priors: from explicit multimodal fusion to inference-time adaptation of generative foundation models"]}]}],"canonical_facts":{"dc:creator":["Tragakis, Athanasios"],"dc:date":["2026"],"dc:date.issued":["2026"],"dc:description.abstract":["Achieving high spatial resolution remains a fundamental challenge in computer vision, where results often fail to adhere to physical constraints. This thesis investigates the use of guided priors to mitigate these limitations, tracing a methodological evolution from explicit multimodal fusion in supervised settings to implicit inference-time priors within generative foundation models across 2D and 3D domains. First, we address depth sensor resolution limits through explicit guidance. We propose IGAF (Incremental Guided Attention Fusion), a deep learning architecture that utilizes high-resolution RGB images to iteratively upsample depth maps. By selectively attending to cross-modal features, IGAF transfers structural guidance from the RGB domain to reconstruct fine geometric details that exceed the sensors native resolution. Shifting toward generative AI, we introduce Pixelsmith, a zero-shot framework for gigapixel image generation on a single GPU. By employing a cascading approach where lower-resolution outputs serve as explicit guidance for patch-based denoising, we leverage the foundation models learned knowledge as an implicit prior. This enables the synthesis of extreme detail while overcoming the memory constraints of standard diffusion models. Finally, we extend this paradigm to 3D with SceneHI, a framework that lifts 2D generative features into 3D space. By orchestrating a rendering / inverse rendering loop as explicit guidance to adapt implicit priors, SceneHI eliminates multi-view inconsistencies. We further integrate geometry-consistent shadows, addressing both resolution and physical plausibility constraints in complex 3D environments. Ultimately, this thesis establishes that structural guidance is an essential requirement for surpassing the physical resolution limits of sensors and foundation models across 2D and 3D domains."],"dc:format":["pdf"],"dc:identifier.doi":["10.5525/gla.thesis.86161"],"dc:identifier.uri":["https://theses.gla.ac.uk/86161/4/2026TragakisPhD.pdf"],"dc:language":["en"],"dc:publisher.institution":["University of Glasgow"],"dc:relation.isreferencedby":["https://theses.gla.ac.uk/86161/"],"dc:subject":["QA76 Computer software"],"dcterms:accessrights":["public"],"dc:title":["Scaling resolution via guided priors: from explicit multimodal fusion to inference-time adaptation of generative foundation models"],"dc:type":["Thesis"],"dc:type.qualificationlevel":["PhD"]},"updated_at":"2026-11-19T08:37:45Z"}