Sensors with high-quality imaging and a 200m-level detection range are essential for Level 4 autonomous driving and beyond. Sensor fusions present immense promise for system robustness to improve driving safety and reliability, while LiDAR stands as a prospective candidate as it offers both high image resolution and depth perception with time-of-flight (ToF). Strong background light contributes the most noise power to >100m LiDAR systems, which drowns signals with conventional transimpedance amplifiers (TIAs) [2]. Pixel accumulating can improve SNR [5] but blurs the image when at a 200m distance. Histogramming [4] enhances the image but deteriorates the frames per second (fps). Both the blurry images and low fps prevent LiDAR from responding to the sudden appearance of vehicles and pedestrians, restricting its capability in autonomous driving. Moreover, reducing the number of LiDARs helps control cost, which nonetheless necessitates a LiDAR capable of both long-range (LR) and short-range (SR) measurements [2]. High-accuracy SR distance measurement is essential for parking assistance, pushing the precision burdens on the TDC and the walk-error compensation. Convolutional neural networks (CNNs) have demonstrated their enormous potential for image resolution upscaling and inpainting [6], particularly as the computing capacity of the vehicles has undergone substantial improvement. In this paper: 1) a high-gain current-integrating-based TIA (CI-TIA) is developed with competitive energy and noise efficiency, which inherits an adaptive background light-compensation feature, thus capable of eliminating the histogramming operations for higher fps; 2) a hybrid pulse position and width converter (PPWC) provides both time and intensity information, while the intensity compensates the walk errors in time intrinsically; the PPWC together with the CI-TIA produce a 50ps (7.5mm) time resolution; and 3) an intensity-guided window-size-adapting accumulation (IGWAA) algorithm improves the accumulation efficiency and accuracy, while a CNN-guided 3D inpainting network further generates missing information. Put together, these result in a 30fps direct ToF LiDAR with a pixel resolution of 256×192 that achieves consistent <1cm distance error and <0.1% precision for up to 240m measurements.