Customizing Video Portraits via Identity-ActionDecoupling

2026-06-21

AI Summary

Existing text-to-video generation methods struggle to simultaneously preserve identity consistency and faithfully follow textual instructions due to facial embeddings entangled with identity-irrelevant information, often resulting in monotonous expressions and distorted motions. To address this, this work proposes an Identity-and-Dynamics decoupling (IaD) framework that, for the first time, injects identity features during inference without requiring subject-specific fine-tuning. The approach explicitly disentangles identity from dynamic attributes through a dedicated identity decoupling loss and a text alignment loss. This enables the generation of videos that not only maintain strong identity consistency but also exhibit rich facial expressions and precise adherence to input prompts, yielding temporally coherent and semantically aligned motion controllable by text.

Abstract

Identity-Preserving Text-to-Video Generation (IPT2V) seeks to synthesize a temporally coherent video from a reference image and a textual description, while simultaneously preserving the subject's identity and allowing fine-grained control over facial dynamics. Although recent methods such as ID-Animator and ConsisID inject identity features only at inference time, they ignored the ID-irrelevant information contained in Facial embedding, leading to monotonous or inaccurate facial movements that poorly follow the prompt. We introduce Identity-Action Decoupling (IaD) framework as well as two loss function Identity Decoupling Loss and Text Alignment Loss to solve this problem. Without any subject-specific fine-tuning, IaD yields videos that (1) maintain cross-temporal identity consistency and (2) exhibit rich, controllable expressions and scene variations that closely match the input text.

Problem

• Identity-Preserving Text-to-Video Generation
• facial dynamics
• temporal coherence
• identity consistency
• text-driven video synthesis

Innovation

• Identity-Action Decoupling
• Identity-Preserving Text-to-Video Generation
• Facial Dynamics Control
• Text Alignment Loss
• Cross-Temporal Identity Consistency