Gespeichert in:
Bibliographische Detailangaben
1. Verfasser: Chadalavada, Hanuma Ramesh
Format: Recurso digital
Sprache:
Veröffentlicht: Zenodo 2026
Online-Zugang:https://doi.org/10.5281/zenodo.19645829
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866901225480388608
author Chadalavada, Hanuma Ramesh
author_facet Chadalavada, Hanuma Ramesh
contents <p>Generative image editing models apply holistic scene-level transformations but systematically degrade semantically structured regions (text, logos, fine patterns) whose fidelity depends on high-frequency detail preservation. We formalize this as context-preserving semantic restoration (CPSR): given an original image and a generatively edited image, recover a result satisfying dual objectives of semantic fidelity on structured regions and context preservation on the scene. General-purpose models (Gemini, FLUX Pro Edit) fail because they lack the domain-specific inverse mapping from artifact space to ground-truth content. We address this via a synthetic degradation curriculum generating paired training data at three severity levels enabling LoRA adaptation (15-25 MB, 1000 steps), a segmentation-guided dual-inference architecture with parallel per-keyword SAM 3 detection, ECC sub-pixel alignment, and feathered mask compositing, and mode-conditioned restoration policies for five editing modalities. Evaluated on 120 images across 8 product lines, our approach achieves 89% text legibility (vs 52% FLUX Pro Edit, 41% Gemini), 0.87 SSIM on device regions, and less than 0.02 LPIPS context deviation. We discuss generalization to automotive, pharmaceutical, retail, and architectural domains.</p>
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_19645829
institution Zenodo
language
publishDate 2026
publisher Zenodo
record_format zenodo
spellingShingle LoRA-Guided Semantic Restoration for Generative Image Editing Artifacts
Chadalavada, Hanuma Ramesh
<p>Generative image editing models apply holistic scene-level transformations but systematically degrade semantically structured regions (text, logos, fine patterns) whose fidelity depends on high-frequency detail preservation. We formalize this as context-preserving semantic restoration (CPSR): given an original image and a generatively edited image, recover a result satisfying dual objectives of semantic fidelity on structured regions and context preservation on the scene. General-purpose models (Gemini, FLUX Pro Edit) fail because they lack the domain-specific inverse mapping from artifact space to ground-truth content. We address this via a synthetic degradation curriculum generating paired training data at three severity levels enabling LoRA adaptation (15-25 MB, 1000 steps), a segmentation-guided dual-inference architecture with parallel per-keyword SAM 3 detection, ECC sub-pixel alignment, and feathered mask compositing, and mode-conditioned restoration policies for five editing modalities. Evaluated on 120 images across 8 product lines, our approach achieves 89% text legibility (vs 52% FLUX Pro Edit, 41% Gemini), 0.87 SSIM on device regions, and less than 0.02 LPIPS context deviation. We discuss generalization to automotive, pharmaceutical, retail, and architectural domains.</p>
title LoRA-Guided Semantic Restoration for Generative Image Editing Artifacts
url https://doi.org/10.5281/zenodo.19645829