Abstract: Objective This study investigates the effect of a 0.3125 mm ultra-high-resolution detector combined with a ClearInfinity (CI) deep-learning reconstruction algorithm on the image quality of orbital computed tomography (CT). Methods Scans were performed using a NeuViz Epoch Elite CT scanner on a Catphan® 600 phantom and three 7-year-old rhesus monkeys. The collimation widths were set to 64 mm ×0.625 mm and 128 mm × 0.3125 mm. Images were reconstructed using filtered back projection (FBP), adaptive iterative reconstruction (ClearView, CV) 60%, and deep-learning reconstruction (CI) 60% algorithms. Image quality was evaluated using objective indicators such as the modulation transfer function (MTF) and contrast-to-noise ratio (CNR), as well as using double-blind subjective scoring. Additionally, statistical analyses were performed. Results In phantom experiments, under standard and bone algorithms, images with a minimum collimation width of 0.3125 mm showed significantly better performances in terms of MTF50%, MTF10%, and some CNR indicators compared with those with a minimum collimation width of 0.625 mm. The CNR of the CI algorithm was significantly higher than those of the FBP and CV algorithms. In animal experiments, the CNR of the medial rectus in images with a 0.3125 mm collimation width was significantly higher than that in images with a 0.625 mm collimation width. The CI algorithm achieved the optimal CNR for the medial rectus and eyeball, as well as the highest subjective scores, with good consistency between two radiologists’ subjective scores (kappa ≥ 0.75). Conclusion The 0.3125 mm ultra-high-resolution detector combined with the CI deep-learning algorithm significantly improved the resolution and contrast of orbital CT images as well as reduced noise and artifacts, thereby demonstrating promising clinical-application prospects.