Existing head avatar animation methods often encounter critical limitations in facial expression transfer, particularly in preserving fine-grained facial details, maintaining high fidelity between generated animations and real facial dynamics, and ensuring precise alignment of expressions with driving signals. These challenges typically result in unnatural animations, loss of identity-specific features, and mismatches between facial movements and input expressions. To address these challenges, we propose DiffusionUNet, a U-Net-based core component of a diffusion model designed for avatar generation and restoration. Specifically, we formulate the tasks of avatar generation and restoration as an inverse problem. DiffusionUNet leverages a symmetric encoder-decoder architecture with skip connections to enable efficient progressive image processing. Furthermore, we introduce a mean loss to balance the pixel-level differences between the images generated by the model and the target images. During the generation and restoration of avatar animations, the facial details of the animations are refined, resulting in the presentation of rich details and high fidelity. Experiments show that our method outperforms existing methods in terms of expression transfer, image detail reconstruction, and expression accuracy during head avatar animation.