kepengxu/realrepv2

This repository contains the official research code for DeCoMix-HDR, a system that converts standard-dynamic-range (SDR) video frames into high-dynamic-range (HDR) output. SDR is the older format used by most cameras and streaming sources, while HDR captures a wider range of brightness and color that modern displays can show. The problem this code tackles is that SDR content comes in many flavors: different cameras and processing pipelines compress highlights, shift colors, or alter saturation in different ways, and most conversion tools assume a single fixed style of SDR input. DeCoMix-HDR is built to handle that variety. The method works in two stages. First, the system learns to recognize the specific type of degradation in an input image by comparing it against other SDR versions of the same scene that were processed differently. It pays separate attention to brightness differences and color differences, because those two kinds of distortion do not always travel together. This comparison step trains the model to build a compact description of what is wrong with the input, without confusing the content of the scene with the quality problem. Second, that description is fed into a reconstruction network that maps the SDR image to HDR. The network applies a broad correction for overall tone and color, then adds a finer spatial pass to recover bright highlights and fix local color artifacts. At inference time, none of the comparison machinery is needed: you give the model one SDR image and it returns the HDR version. The code is structured around the BasicSR training framework, which is a standard tool in academic image restoration research. Setup requires Python 3.8 or later, PyTorch with a CUDA-capable GPU, and a handful of additional libraries listed in requirements.txt. The training data is the HDRTV4K and HDRTV1K datasets, which must be downloaded separately from the HDRTVDM repository and converted to LMDB format before training can start. This is academic research code released alongside a paper that was not yet publicly linked at the time of this release. There are no pre-trained weights bundled in the repository, so running evaluation requires completing the full training run first, which the config sets at 400,000 iterations. The repository is best suited for researchers or engineers working on video quality and HDR conversion who want to reproduce or build on this specific method.