The growing demand for autonomous rendezvous, inspection, and Active Debris Removal (ADR) missions calls for reliable vision-based relative navigation under challenging orbital conditions, where classical feature-based pipelines often fail. Deep learning provides a powerful alternative, as modern CNN/Transformer backbones can extract robust, illumination-tolerant features even under texture poverty, occlusions, or degenerate views. This work benchmarks three representative AI-based 6 Degrees of Freedom (6-DoF) pose estimation approaches for non-cooperative spacecraft: YOLOv8+SQPnP, the new YOLOv11+SQPnP, and NVIDIA CenterPose. A custom Blender-generated dataset of the Deimos-1 satellite comprising 16,200 photorealistic grayscale images with systematically varied viewing and illumination geometries was used for training and evaluation. Results in controlled fly-around scenarios show that YOLOv11+SQPnP achieves the best overall balance between geometric accuracy and temporal stability, with translation errors around 45–60 mm and orientation errors near 2–3◦ across most viewpoints. CenterPose DLA-34 remains the most robust under adverse illumination or self-shadowing, consistently maintaining low variance and smooth trajectories (translation errors ∼52–55 mm, rotation errors ∼3.5◦ ). YOLOv8+SQPnP provides the sharpest geometric fits (3DIoU up to 0.92) but suffers from high sensitivity to viewpoint and lighting. Within this controlled synthetic benchmark, these findings indicate that compact detector-plus-PnP pipelines can offer a lower-complexity alternative to heavier integrated backbones and are promising candidates for future onboard assessment, although embedded suitability was not benchmarked here.