1. Trained on a higher-resolution dataset

2. The training set contains no faces; theoretically, high weights will not affect face generation

3. Adjusted training parameters

4. Improved texture quality

Reprinted from: z-image Turbo Detail Enhancement v2