Answers are grounded in this paper and cite their sources.
Portrait images often suffer from poor colour, low contrast, and improper exposure. Ultra-high-resolution enhancement relies on manual adjustments, while deep-learning methods are limited by large parameter requirements and insufficient high-quality training samples. This study proposes an end-to-end ultra-high-resolution portrait enhancement method using image signal processing (ISP) curve prediction, with a global ISP branch and a local curve branch. The global branch uses cross-attention for global modelling and predict a colour correction matrix and gamma parameters. The local branch extracts low-resolution features using a lightweight convolutional residual structure and predicts spatially varying residual mappings parameterized by RGB values and spatial coordinates for full-resolution enhancement. Extensive comparative experiments were conducted on the ALSPP5 K multi-scene portrait, MIT-Adobe FiveK, and HDR + datasets. On MIT-Adobe FiveK, the proposed method achieves 25.67 dB PSNR, 0.960 SSIM, and 0.032 LPIPS. Compared with Star-Enhancer, this corresponds to a 0.23 dB increase in PSNR, a 0.012 increase in SSIM, and a reduction in LPIPS from 0.083 to 0.032 (approximately 61.4%).
No takes yet. Share an insight, caveat, or question.
Zheng et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: