Fig. 4: Multi-level consistent performance evaluation of EyeReal and passive SBP-utilization models in large-scale displays.

a, Local-scale spatial performance comparison of EyeReal and modern view-segmented automultiscopy across binocular surroundings. b, Cross-scene and cross-pose generalization evaluation of EyeReal. It maintains robustness with high-quality rendering on previously unseen scenes and adaptability across a wide range of new head poses. Error bars represent the standard error of the mean with 6,000 samples. c, Global-scale spatial performance comparison of EyeReal and the iterative view-dense representative based on NTF with up to 50 iterations (iter.) across different visual-field ranges. NTF beyond 50 iterations was excluded because of speeds below 1 Hz (see e). Error bars show the standard deviation. d, Global-scale spatial performance comparison of EyeReal and the NVD representative⁸. As neural methods are trained on fixed views under predefined distances and directions, we compare multiple distance–direction combinations to highlight the differences. Error bars denote standard deviation. e, Runtime comparison of EyeReal, NTF and NVD within a sub-second-level timeframe. Except for the performance in b and c, EyeReal achieves real-time capability that is one to two orders of magnitude faster. f, Focal discrimination curves across uniformly sampled depths under the same aperture, from foreground to background. Each region shows smooth transitions in clarity with a peak at its corresponding depth. F–M region, a specific region between the foreground and midground; M–B region, a specific region between the midground and background. g, PSNR point-cloud heatmap across the visual field. Each point represents a randomly sampled viewpoint, with around 600,000 samples in total. Sampling at larger radial distances is omitted because of consistently high PSNR in those regions. PSNR, peak signal-to-noise ratio; SSIM, structural similarity index measure.