HDR Exposure Synthesis

A pipeline that tries to reconstruct HDR from one LDR image by synthesising two extra exposures. The dataset loader reads exposure times from EXIF, sorts the bracketed frames and picks the ones closest to +-2.7 EV relative to the middle frame. A U-Net with a six-channel output predicts both exposures at once, evaluation runs tiled full-resolution inference, merges the three images with the Debevec algorithm and compares the dynamic range with the reference HDR. Only 16 scenes had usable EXIF for training and 3 for testing, the model overfitted, PSNR of the generated exposures was 15.8 dB (under) and 18.6 dB (over), and the merged HDR covered about 4 EV against 19-25 EV for the references. The repository documents these limitations and the next step (a smaller model with dropout).