Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Identity Preserving 3D Head Stylization with Multiview Score Distillation

Created by
  • Haebom

Author

Bahri Batuhan Bilecen, Ahmet Berke Gokmen, Furkan Guzelant, Aysegul Dundar

Outline

This paper presents a novel framework for 3D head styling that addresses the challenges of existing methods, which primarily rely on frontal photos, while maintaining individuality. We synthesize 360-degree field-of-view images using the PanoHead model, and integrate negative log-likelihood distillation (LD), multi-view grid scores, mirror gradients, and score rank weighting techniques into a 3D generative adversarial network (GAN) architecture to enhance individuality preservation and styling quality. This provides insight into the effective distillation process between diffusion models and GANs, with a particular focus on individuality preservation.

Takeaways, Limitations

•
Takeaways:
◦
Expand styling possibilities from various angles by utilizing the 360-degree field of view.
◦
Preserving personality and improving styling quality through negative log-likelihood distillation (LD).
◦
Improved 3D GAN architecture using multi-view grid scores and mirror gradients.
◦
Provides new insights into the effective distillation process between diffusion models and GANs.
•
Limitations:
◦
Since the structure is dependent on the PanoHead model, there is a possibility of performance degradation when applying other 3D models.
◦
Further validation of the generalization performance of the proposed method is needed.
◦
Lack of performance evaluation on large datasets.
👍
Made with Slashpage