Authenticity-constrained generative image enhancement using semantic consistency and provenance metadata

Abstract The use of digital image enhancement has advanced with recent developments in artificial intelligence (AI), resulting in image enhancements that are visually pleasing and can sometimes exceed the capabilities of traditional methods. In digital media systems, however, a number of current deep learning approaches focus on perceptual realism, rather than semantic authenticity, which can lead to contextual distortion, misinformation risk, and lack of traceability of the provenance of the media product. To address this limitation, we introduce an Authenticity-Constrained Generative Model (ACGM) that jointly balances perceptual enhancement, semantic authenticity, and provenance traceability. The architecture combines a U-Net-based generator with residual dense blocks, a CLIP/VGG-based authenticity-preservation module, and a provenance-aware traceability encoder. These components are jointly optimized using the Creativity–Authenticity Loss Function (CALF), which integrates adversarial, perceptual, authenticity, and traceability terms. Semantic authenticity is assessed through input–output feature consistency, whereas traceability is evaluated by embedding and recovering invisible, C2PA-inspired provenance metadata. The present implementation does not generate a formally signed C2PA manifest. The development set consisted of 1,500 paired samples from Flickr2K and MIT-Adobe FiveK, of which 1,050 were used for training, 225 pairs were used for validation, and 225 pairs were used for internal testing. Under the reported controlled conditions, ACGM obtained PSNR = 33.7 ± 0.58, SSIM = 0.937 ± 0.009, LPIPS = 0.071 ± 0.007, FID = 10.5, and CLIP-based semantic similarity = 0.91 ± 0.02. Overall, the results of human-centered evaluation showed that there were significant model effects within participants for perceived visual appeal, perceived authenticity, and trustworthiness (all p<0.001). Bonferroni-adjusted paired comparisons revealed that ACGM scored significantly higher than all the evaluated baselines on all three criteria. These findings remain preliminary because of the modest dataset scale, the controlled provenance-recovery experiment, and the absence of robustness testing under common image transformations.