UniID: Training for Identity, Inference for Controllability Official model weights for “Training for Identity, Inference for Controllability: A Unified Approach to Tuning-Free Face Personalization”.
Model Description
UniID is a unified tuning-free framework for face personalization that synergistically combines text embedding and adapter approaches. It is designed to achieve high identity fidelity while preserving the text controllability of the original diffusion model.
The key innovation lies in a strategic training-inference paradigm:
Training: An identity-focused learning scheme guides both the text embedding and adapter branches to capture exclusively identity-relevant features.
Inference: A normalized rescaling strategy recovers the text controllability of the original diffusion model and allows the identity signals from both branches to mutually reinforce each other.
This model repository contains the pre-trained weights for both the text_branch and the adapter_branch.
Model Architecture
UniID consists of two main components that work together:
Text Branch: Modifies the text embedding space to incorporate the target identity.
Adapter Branch: Uses an adapter (similar to IP-Adapter) to inject identity features from a reference image into the diffusion model.
For inference, both branches must be loaded and used simultaneously.
More Information
For the full inference and training code, detailed documentation, and more examples, please visit our GitHub repository:
Citation
If you use this model, please cite our work:
@article{UniID,
title = {Training for Identity, Inference for Controllability: A Unified Approach to Tuning-Free Face Personalization},
author = {Lianyu Pang, Ji Zhou, Qiping Wang, Baoquan Zhao, Zhenguo Yang, Qing Li, Xudong Mao},
journal = {arXiv preprint arXiv:2512.03964},
year = {2025}
}
- Downloads last month
- -