Layout, Geometry or Generator? Dissecting Two-Stage Satellite-to-Street View Synthesis with Frozen Foundation Features

Satellite-to-street view synthesis asks a model to imagine a ground-level panorama from a single overhead tile. We dissect a two-stage pipeline on CVUSA, in which a semantic layout predicted from the tile guides a conditional generator, and measure which design choices actually matter. Stage 1, a lightweight PolarLayoutNet (1.53M trainable parameters) on frozen DINOv2 features, reaches 0.514 mIoU against SegFormer pseudo-labels, +0.112 over a CrossNet retrained on the same labels. The backbone (+0.028 over VGG-16) and full-tile input (+0.012 over a 224px crop) drive this gain, whereas the polar mapping, ground-plane prior and learned offsets change mIoU by at most 0.001. In Stage 2, conditioning SPADE on a geometrically warped aerial image consistently helps: LPIPS falls for 87.0% of test images with oracle layouts and 85.3% with predicted ones, with paired 95% bootstrap intervals excluding zero. With oracle layouts, clean-FID drops from 167.4 for a Pix2Pix baseline to 141.0 (116.2 under our earlier development harness, which is not directly comparable); end-to-end with predicted layouts it is 182.1. This oracle-to-predicted gap is the largest effect we measure and identifies layout quality as the main lever for improvement. Published methods remain ahead (FID 24–41 under different protocols), and our labels are pseudo-labels from one dataset and one training seed. The study used 47.8 logged GPU-hours; code, logs, checkpoints and the evaluation harness are released. Code, logs, and checkpoints are fully open-source and available at: https://github.com/anasahmed81103/AerialToGround