The dataset used for training consisted of images from a single environment with limited sunlight. As a result, the generated images tend to lack significant sunlight. By reducing the LoRA strength (if necessary), you can achieve better and more balanced results.