BabyFlow: 3D modeling of realistic and expressive infant faces

TL;DR

BabyFlow introduces a flow-based, disentangled model of infant facial shape and expression that copes with limited data and spontaneous expressions by using normalizing flows for identity and expression latents. It enables cross-age expression transfer from adults to infants, improves 3D reconstruction in expressive regions, and supports controllable 3D synthesis and diffusion-based 2D image generation with geometry-consistent outputs. The approach yields well-behaved latent spaces and demonstrates practical data-augmentation benefits, while uncovering limitations such as dataset size and the need for clinical validation. Overall, BabyFlow advances infant craniofacial analysis by combining cross-age data augmentation, explicit identity-expression disentanglement, and cross-modal synthesis for robust downstream tasks.

Abstract

Early detection of developmental disorders can be aided by analyzing infant craniofacial morphology, but modeling infant faces is challenging due to limited data and frequent spontaneous expressions. We introduce BabyFlow, a generative AI model that disentangles facial identity and expression, enabling independent control over both. Using normalizing flows, BabyFlow learns flexible, probabilistic representations that capture the complex, non-linear variability of expressive infant faces without restrictive linear assumptions. To address scarce and uncontrolled expressive data, we perform cross-age expression transfer, adapting expressions from adult 3D scans to enrich infant datasets with realistic and systematic expressive variants. As a result, BabyFlow improves 3D reconstruction accuracy, particularly in highly expressive regions such as the mouth, eyes, and nose, and supports synthesis and modification of infant expressions while preserving identity. Additionally, by integrating with diffusion models, BabyFlow generates high-fidelity 2D infant images with consistent 3D geometry, providing powerful tools for data augmentation and early facial analysis.

Figures (13)

• Figure 1: Overview of the BabyFlow's pipeline: 1) estimate a mapping $\mathbf{\mathscr{M}}$ that uses barycentric coordinates to align the adult FaceWarehouse Zhou2014 meshes with the BabyFM Morales2025 mesh topology; 2) detect facial expressions by recognizing action units (AUs) Ekman2002 using a spectral representation of 3D geometry; 3) transfer a full set of expressions from the adult surfaces to the infant meshes; 4) apply a tensor factorization technique, to separate identity and expression components, resulting in a bilinear model; 5) use normalizing flows to model and sample from the latent spaces of identity and expression, enabling the generation of realistic and expressive infant faces.
• Figure 2: Cross-age expression transfer from adults to infants. The first column shows a baby an infant identity (bottom row) and an adult identity (top row), both displaying a neutral expression. In the remaining columns, the top row presents various target expressions of the adult subject, and the bottom row shows the corresponding expressions transferred from the adult to the infant.
• Figure 3: Mean error per vertex across the test set, obtained by fitting the 3D scans using the BabyFlow (left) and the BabyFM Morales2025 (right). Warmer colors correspond to higher errors.
• Figure 4: Examples of 3D synthetic infant faces generated with BabyFlow by randomly sampling identities and expressions.
• Figure 5: Latent space interpolation of the normalizing flow models. The first row varies only the expression, the second varies only the identity, and the third varies both identity and expression simultaneously. Each column represents an intermediate step in the transition, with interpolation increments of $0.25$.