Experiments

Visual Comparisons

We compare our approach with other enhancement methods. SuperHead synthesizes high-quality facial details across diverse expressions, clearly outperforming baselines and in some cases approaching the pseudo ground-truth head avatar. All methods are driven and rendered with novel camera poses and expressions.

Quantitative Results

State-of-the-art Performance: We evaluated SuperHead on the NeRSemble and INSTA datasets. Our method consistently achieves the best scores across key image quality metrics:

• PSNR / SSIM ($\uparrow$): Higher values indicate better fidelity. SuperHead outperforms all baselines.
• LPIPS ($\downarrow$): Lower values indicate better perceptual quality. Our method produces the most natural-looking results.
• Efficient Processing: Thanks to our efficient 3D GAN inversion, SuperHead requires significantly less inference time compared to video-based super-resolution methods (e.g., 5 mins vs. 30 mins for a sequence), making it a practical solution for scalable avatar creation.

For more video comparisons, see the SuperHead website for more information.