Latent-Identity Tuning in Text-to-Image Personalization Models

TL;DR

The paper introduces a technique to make precise, localized edits to a person's identity in text-to-image models by manipulating latent representations in a frozen encoder. It identifies semantic directions in the latent space tied to specific facial regions without retraining. The method enables consistent identity generation across diverse images with fine-grained control over facial features.

Why It Matters

This approach allows for accurate personalization of facial images in generative models, which is crucial for applications like digital avatars, virtual identity editing, and personalized content creation. It achieves high fidelity without requiring model retraining, making it efficient and practical for real-world deployment.

Key Findings

• Latent space of a frozen text-to-image encoder contains identifiable semantic directions corresponding to facial regions.
• Selective manipulation of specific latent tokens enables localized, semantically meaningful facial edits while preserving identity consistency.
• The method achieves diverse, high-quality image generation with consistent identity across variations, using only inference-time adjustments.