The creation of realistic avatars in motion is a hot-topic in academia and the creative industries. Recent advances in deep learning and implicit representations have opened new avenues of research, especially to enhance the details of the avatars using implicit methods based on neural networks. State-of-the-art implicit Fast-SNARF methods encodes various poses of a given identity, but are specialized for that single identity. This paper proposes NIPIG, a method that extends Fast-SNARF model to handle multiple and novel identities. Our main contribution is to condition the model on identity and pose features, such as an identity code, a gender indicator, and a weight estimate. Extensive experiments led us to a compact model capable of interpolating and extrapolating between training identities. We test several conditioning techniques and network's sizes to find the best trade-off between parameter count and result quality. We also propose an efficient fine-tuning approach to handle new out-of-distribution identities, while avoiding decreasing the reconstruction performance for in-distribution identities.
Abstract
Teaser Video
Interactive Controls
Method Overview
NIPIG extends Fast-SNARF with forward skinning in canonical space and an occupancy network conditioned on identity and pose. Given a posed query point, we iteratively find canonical correspondences via the LBS field (conditioned on identity), evaluate occupancy (conditioned on identity and pose), and reconstruct the surface. The model can also be conditioned on other attributes such as a weight or gender parameter.
Conditioning Techniques
We review three standard mechanisms to condition a layer \(l\) with a code \(c \in \mathbb{R}^{n_c}\). Let \(x^{(l)} \in \mathbb{R}^{n_l}\) be the input activation and \(\sigma\) a nonlinearity. The output is \[ x^{(l+1)} = f^{(l)}(x) \quad \text{with} \quad f^{(l)}(x) = \sigma\!\big(W\, x + B\big), \]
- Concatenation. Augment the layer input with the conditioning code: \(\tilde{x}^{(l)} = [\,x^{(l)}; c\,] \in \mathbb{R}^{\,n_l + n_c}\). The layer becomes \[ x^{(l+1)} = f^{(l)}(\tilde{x}^{(l)}) = \sigma\!\big(W\,\tilde{x}^{(l)} + B\big), \quad W \in \mathbb{R}^{\,n_{l+1}\times (n_l + n_c)},\; B \in \mathbb{R}^{\,n_{l+1}}. \]
- FiLM (Feature‑wise Linear Modulation). Two learned projections of \(c\) produce a per‑feature gain and bias, \(\gamma(c), \beta(c) \in \mathbb{R}^{\,n_{l}}\). The output of layer \(l\) becomes: \[ x^{(l+1)} = f^{(l)}\big(\gamma(c) \odot x^{(l)} + \beta(c)\big), \] where \(\odot\) denotes element‑wise multiplication.
- Hypernetworks. A hypernetwork \(H^{(l)}\) maps the code to the actual layer parameters, \(H^{(l)}(c) = \{\,W^{(l)}(c),\, B^{(l)}(c)\,\}\), with \(W^{(l)}(c) \in \mathbb{R}^{\,n_{l+1}\times n_l}\) and \(B^{(l)}(c) \in \mathbb{R}^{\,n_{l+1}}\). The conditioned layer is \[ f^{(l)}_c(x^{(l)}) = \sigma\!\big(W^{(l)}(c)\,x^{(l)} + B^{(l)}(c)\big). \] This approach often yields strong performance but typically increases parameter count and memory usage compared with concatenation or FiLM.
Teacher-Student Fine-Tuning
To add a new out-of-distribution identity without catastrophic forgetting, we duplicate the model into a frozen teacher and a trainable student. The student is supervised by ground truth for the new identity and by the teacher's predictions for prior identities. This allows fast adaptation (≈1k iterations) while preserving performance on the original identities.
Pose Generalization in Extreme Motions
NIPIG reposes identities robustly on sequences far from the training distribution (e.g., MPI PosePrior). The forward-skinning field avoids the instability of inverse-skinning approaches and yields clean surfaces even for new identities (see image below).
Comparison to COAP and Fast-SNARF
Qualitatively, NIPIG reduces part-boundary seams typical of compositional models like COAP and produces sharper facial details than Fast-SNARF—particularly due to our higher-density sampling near the face. The model is as compact as Fast-SNARF while representing many more identities.
Quantitative Results & Model Size
Across seen and unseen settings, NIPIG matches or surpasses Fast-SNARF and COAP on IoU and Chamfer while using substantially fewer parameters. The neutral-gender single model also performs strongly with a footprint ∼0.52M params (base) or 0.13M (light).
Fine-Tuning Results & Forgetting
Supervised teacher-student fine-tuning mitigates forgetting on previously seen identities while retaining quality on the new identity. In both cases, the fine-tuning converges in a few hundred iterations.
BibTeX
@inproceedings{nipig_cvmp25,
title={ NIPIG: Neural Implicit Avatar Conditioned on Human Pose, Identity and Gender },
author={ Guillaume Loranchet and Pierre Hellier and Adnane Boukhayma and João Regateiro and Franck Multon },
booktitle={CVMP 2025},
year={2025}
}