Information-Regularized Constrained Inversion for Stable Avatar Editing from Sparse Supervision

2026-04-03

AI Summary

This work addresses the ill-posed inversion problem in animatable portrait editing under sparse supervision, which often leads to identity leakage and temporal flickering due to insufficient constraints. To mitigate these issues, the authors propose a constraint inversion framework grounded in information-theoretic regularization. The approach restricts edits to low-dimensional, part-specific subspaces within a structured latent space and derives an information matrix via local linearization of the decode-and-render pipeline. The spectral properties of this matrix are leveraged to predict edit stability, guiding frame reweighting and keyframe activation. By integrating efficient Hessian-vector product computation with conditional objective optimization, the method significantly enhances temporal consistency while effectively suppressing identity leakage and flickering artifacts, even under limited supervision.

Abstract

Editing animatable human avatars typically relies on sparse supervision, often a few edited keyframes, yet naively fitting a reconstructed avatar to these edits frequently causes identity leakage and pose-dependent temporal flicker. We argue that these failures are best understood as an ill-conditioned inversion: the available edited constraints do not sufficiently determine the latent directions responsible for the intended edit. We propose a conditioning-guided edited reconstruction framework that performs editing as a constrained inversion in a structured avatar latent space, restricting updates to a low-dimensional, part-specific edit subspace to prevent unintended identity changes. Crucially, we design the editing constraints during inversion by optimizing a conditioning objective derived from a local linearization of the full decoding-and-rendering pipeline, yielding an edit-subspace information matrix whose spectrum predicts stability and drives frame reweighting / keyframe activation. The resulting method operates on small subspace matrices and can be implemented efficiently (e.g., via Hessian-vector products), and improves stability under limited edited supervision.

Problem

• avatar editing
• sparse supervision
• identity leakage
• temporal flicker
• constrained inversion

Innovation

• constrained inversion
• avatar editing
• information regularization
• edit subspace
• sparse supervision