UniLumos is designed to address the limitations of existing video relighting models, which often produce unrealistic results due to the lack of physical correctness in visual space. By supervising the model with depth and normal maps extracted from its outputs, UniLumos explicitly aligns lighting effects with the scene structure, enhancing physical plausibility. The framework also employs path consistency learning, allowing supervision to remain effective even under few-step training regimes. This approach enables UniLumos to achieve state-of-the-art relighting quality with significantly improved physical consistency.
Lumos-Custom also introduces LumosBench, a structured benchmark that targets six core illumination attributes defined in the annotation protocol. The benchmark evaluates the performance of video relighting models from two key perspectives: qualitative alignment with user prompts and quantitative measurement of physical lighting properties. This benchmark enables automatic and interpretable assessment of relighting precision across individual dimensions, providing a comprehensive evaluation of the framework's performance. The project also provides a set of tools and scripts for easy installation and usage, making it accessible to researchers and developers.
