New semi-realistic images were incorporated for training the 2.5 version model. The model's generation capabilities were significantly improved through post-training on the 2.0 version using Direct Preference Optimization (DPO), resulting in outputs that more closely align with human preferences.