Abstract
High dynamic range (HDR) imagery provides a rich representation of scene radiance, but remains challenging for diffusion models trained on bounded, perceptually compressed imagery. A natural approach is to learn a mapping from HDR data into the latent space of a pretrained diffusion model. However, this requires large HDR datasets and substantial additional training. In this work, we present a framework for SDR-to-HDR video translation and text-to-HDR video generation, leveraging the visual priors of pretrained diffusion models. We observe that applying a logarithmic encoding, commonly used in cinematic pipelines, to HDR videos produces representations that are naturally aligned with the latent space of these models. This alignment enables adapting pretrained diffusion models for HDR generation through lightweight fine-tuning, without modifying the latent space in which they operate or requiring an explicit HDR-to-latent mapping. To encourage the model to infer missing HDR content from its learned priors, we augment SDR-to-HDR training with camera-mimicking degradations that require recovering lost details. Using only lightweight adaptation of a pretrained video diffusion model, we demonstrate high-quality HDR video generation from both text and SDR video across diverse scenes and challenging lighting conditions. Our results show that HDR can be effectively modeled when its representation is aligned with the model's learned priors.
