We propose a white-balanced HDR environment map dataset to evaluate the color robustness of lighting estimation models. We demonstrate that preprocessing inputs with a pre-trained white balance network significantly improves color accuracy without requiring retraining.