HDR Video Generation via Latent Alignment with Logarithmic Encoding
This work generates high-quality HDR video from standard SDR input by aligning its data distribution with a pretrained generative model's latent space. The method uses a LogC3 encoding to map scene-linear radiance into the model's expected range, treating HDR content as familiar SDR data. Targeted training using camera degradations helps the model infer missing details, resulting in temporally coherent HDR video without retraining the core encoder.