# BlendFields: Few-Shot Example-Driven Facial Modeling

**Authors:** Kacper Kania, Stephan J. Garbin, Andrea Tagliasacchi, Virginia Estellers, Kwang Moo Yi, Julien Valentin, Tomasz Trzciński, Marek Kowalski  
**arXiv:** [2305.07514](https://arxiv.org/abs/2305.07514)  
**Date:** 2023-05-12  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.07514v1/

## Key points

- Accurate visualizations of human faces require both coarse and fine-level details of facial geometry and appearance
- Existing methods either rely on data-driven approaches or fail to capture fine details due to reliance on geometric face models
- BlendFields is a new method that bridges this gap by drawing inspiration from traditional computer graphics techniques
- BlendFields extends VolTeMorph to enable modeling of high-frequency information such as expression wrinkles in a few-shot setting
- Correctives to neural field representations are introduced and activated according to local deformations, making facial modeling more accessible
- The model generalizes beyond facial modeling, showing its effectiveness in modeling wrinkles on a deformable object made of rubber
- BlendFields finds a sweet spot between volumetric representations and explicit models by using limited amounts of data to learn missing details in controlled models and combining them together
- BlendFields overcomes limitations of Neural Radiance Fields (NeRF) by using a limited amount of data while still being able to learn fine-grained effects on top of smooth volumetric deformations
- Overall, BlendFields is an effective method for generating faithful visualizations of human faces that captures both coarse and fine-level details while also being applicable beyond facial modeling.

## Summary

The generation of accurate visualizations of human faces requires the capture of both coarse and fine-level details of facial geometry and appearance. However, existing methods either rely on data-driven approaches that require extensive datasets not accessible to the research community or fail to capture fine details due to reliance on geometric face models that cannot represent texture with mesh discretization and linear deformation designed for modeling only coarse face geometry. In this paper, the authors introduce a method called BlendFields that bridges this gap by drawing inspiration from traditional computer graphics techniques. BlendFields extends VolTeMorph to enable modeling of high-frequency information such as expression wrinkles in a few-shot setting. The authors also introduce correctives to neural field representations and activate them according to local deformations. This approach makes facial modeling more accessible by providing an alternative to data and compute-intensive techniques. The authors demonstrate how their model generalizes beyond facial modeling, showing its effectiveness in modeling wrinkles on a deformable object made of rubber. In comparison with related works, BlendFields finds a sweet spot between volumetric representations and explicit models by using limited amounts of data to learn missing details in controlled models and combining them together. While Neural Radiance Fields (NeRF) have grown in popularity for their ability to represent complex geometries accurately, they are constrained to static scenes and lack ways to control the scene. BlendFields overcomes these limitations by using a limited amount of data while still being able to learn fine-grained effects on top of smooth volumetric deformations. Overall, BlendFields is an effective method for generating faithful visualizations of human faces that captures both coarse and fine-level details while also being applicable beyond facial modeling. The authors' empirical evaluations confirm the advantages of their approach compared with related works, making it an exciting contribution towards advancing computer graphics techniques.