Identity-Consistent Expression Fields: A Disentangled NeRF Framework for Few-Shot Facial Expression Synthesis

The paper proposes Identity-Consistent Expression Fields (ICEF), a few-shot facial-expression synthesis framework built on disentangled neural radiance fields. It separates a static, identity-specific radiance component from an expression-conditioned deformation component, then applies an identity-preservation regularizer to restrict changes to expression-relevant regions. A confidence-weighted feature-warping step reduces the influence of unreliable warps when target expressions are far from the observed few-shot inputs. The abstract describes an evaluation protocol for rendering quality and identity consistency across extrapolation distances, but reports no quantitative results.